FYI: It works fine on my Intel NUC (Haswell i3 Ultra Low TDP)

olejon on 8 Jun 2020

Hi
Can you check Kernel log if corefreqk.ko has crashed ?

EDIT: Forget log. Here is the reason.

I've split the driver shared memory pages into two groups. Those with read-only data, the others with read-write data.
When Daemon is violating access (writing into a RO page) it will segfault.

I have developed and tested with Intel, but not with AMD. Sorry for this mess.

I will come back with a fix.

Also, by the end of the week, my new 3950X should be ready for the AMD developments

cyring on 9 Jun 2020

👍1

Well, it stays loaded, and the corefreqd-cmgr process must be killed to unload it.

The kernel message is as it used to be IIRC. As it's not signed and loaded with insmod, although Secure Boot is disabled, but boots using EFI, not legacy.

corefreqk: loading out-of-tree module taints kernel.
corefreqk: module verification failed: signature and/or required key missing - tainting kernel
CoreFreq(1:7): Processor [ 8F_71] Architecture [Zen2/Matisse] SMT [12/12]

However the NUC, also using EFI boot partition, only shows:

CoreFreq(0:2): Processor [ 06_45] Architecture [Haswell/Ultra Low TDP] SMT [4/4]

But the X570 has a more modern BIOS for (U)EFI. The NUC is running the cleanest netinstall of Ubuntu 18.04. There are som differences there.

olejon on 9 Jun 2020

I'm searching in source code for a change concerning an AMD instruction or RO page addressing but I don't find any so far.
It will be helpful if you start the Daemon with gdb and tell me in which function it segfaults

Build _CoreFreq_ with debug flag

make clean all CC='cc -g'

Load the driver and start Daemon in gdb

gdb corefreqd
run

Please post the output of GDB when it crashes and its call stack

backtrace

Thanks

cyring on 9 Jun 2020

OK, I guess I found the issue.
Forget my gdb request above and just pull and try the last commit
Thanks

cyring on 9 Jun 2020

Indeed. It works :+1:

I still compile like this:

make MSR_CORE_PERF_UCC=MSR_IA32_APERF MSR_CORE_PERF_URC=MSR_IA32_MPERF

Cheers!

olejon on 9 Jun 2020

Great. Thanks for your test.

cyring on 10 Jun 2020

Also tested on Ubuntu 20.04 LTS with Live USB. Works fine there too :+1:

olejon on 10 Jun 2020

Also tested on Ubuntu 20.04 LTS with Live USB. Works fine there too

Thanks a lot

cyring on 10 Jun 2020

Also, by the end of the week, my new 3950X should be ready for the AMD developments

Cool! Finally ordered mine as well, with this cooler. What a beast! More than 1.3 kg... Only worry it'll make a nice part of the inside invisible through the tinted chassis window.

https://noctua.at/en/nh-d15-se-am4

olejon on 11 Jun 2020

Freshly setup

I'll be out of coding for a while. There's some fans to setup, many tunings, close the case, and my GPU is still my old GTX570. etc
But so far those G.Skill DIMM die Samsung are amazing : 3600 MHz CL 16 with no effort
Fractal is making great products : Define 7 TG grey case + Celsius Dynamic S36. Temps looks goods.
Best for the end: Ryzen 3950X detected by _CoreFreq_

cyring on 11 Jun 2020

Nice! :+1:

If your RAM chips are 3600 MHz, IIRC from this video tuning Infinity Fabric to not be 50/50 shows benefit. My 3200 MHz ones don't. Also other tweaks, although little overclocking can be done on Ryzen, because it shouldn't be necessary... Intel sells on assuming people will do it. https://youtu.be/7w1EGPZUESU

olejon on 11 Jun 2020

Thanks for the tips. Just gave a quick look but the video games intro makes me think about getting a new GPU which is Linux friendly. AMD RX 5600 is in my short list. But I'm not a gamer at all, and Pro GPU, such as the AMD WS series PCI 4.0 is too expensive, just to run a Linux desktop !
Really, manufacturers are totally ignoring that professionals need some thing else than RGB & gaming components

cyring on 11 Jun 2020

Yes. Well they use like Cinebench etc a lot as it is a good benchmark tool. Games shows an overlay with tons of info including RAM performance/usage, GPU memory performance/usage if using MSI Afterburner (free and not limited to MSI cards nor AMD cards, works on Intel and NVIDIS GPUs too).

That's why I stick to my Radeon RX 570 (MSI), which was on discount and everything. It's still sold here, but the MSI one seems hard to find now.

Chose it after considering an NVIDIA 1660 as they've always been touted as very good vs price, and especially because almost always big discount on many sites. But didn't really want NVIDIA. Saw the Norwegian review and was compared to the RX 570. For me most important was of course out of the box Linux support (in kernel), then noise and performance at idle were basically the same (that review used Windows of course).

It gives 300 FPS using a FPS-testing map in CS:GO so... Why switch? Way above my monitors' refresh rates.

GPUs nowadays are the biggest single investment and they know it ( unless buying a 3950X :P ). They Ramp up prices because little competition. AMD tries to stay just below NVIDIA, that's about it. But at least their new 5XXX series is supported by the open radeon driver.

olejon on 11 Jun 2020

Does Linux reclocking work with the RX 570 ?
The RX 5000 series brings PCIe 4. Not sure it makes a performance difference with previous generation. I'll avoid RX 5500 because of its 8x speed, although it's the cheapest.
You may have read that kernel changelogs leak the ID of next Radeon GPU (EDIT: Sienna Cichlid) . Could be worth to wait for it. But I don't expect a cheap price, day one.
NVIDIA is making the fastest GPU. I have been satisfied with its closed source until version 390xx got unsupported by Arch Linux. The only AUR package for 390xx is most of the time in late with latest Linux kernel.
Nvidia should really contribute to the nouveau driver.

cyring on 11 Jun 2020

Haven't tried, haven't even searched so don't really know how to (you do?). Pretty sure I tried the max OC which MSI Afterburner supports with a click, on Windows, and it's safe, never void warranty, and it worked just fine

BIOS says it runs x16 Native according to, notice the OC in the name, Specifications of Radeon RX 570 ARMOR 8G OC

Yeah, maybe. Haven't followed tech for a while. I'm happy as long as the desktop is super smooth, HW video acceleration is smooth, Wayland works, Android Emulator is smooth, games are smooth although I don't game but have "played around a bit", online and with bots when last lived in Spain
NVIDIA tends to be winning. But for me going from Intel iGPU to a dedicated AMD GPU was nice, TBH. Having 3 DisplayPorts and 1 HDMI is nice. Though NVIDIAs Control Center allows for some crazy setup xrandr simply can't do, although those changes can be written to xorg.conf. I use a pretty crazy multi-monitor setup on my monster laptop at my parents house (laptop 18 inches, VGA primary 17 inches + HDMI TV)
Yeah, when I changed to Ryzen and realized it had no iGPU I looked first at NVIDIA. Then entered the world of AMD having a great driver built-in, and they actively contribute. I've seen improvements after kernel updates and graphics stack updates. I looked at the nouveau supported cards. None are really supported. They all lacked something I want. Was like "yay it supports X" then footnote "doesn't support Y, or is buggy". IIRC saw the Arch Wiki for helping me choose but basically said none have what one should expect today. And the performance sucks even if something is supported
You won't be very impressed by the results of say Unigine Superposition benchmark for a RX 570. It performes a bit better on Windows where it uses DirectX, so maybe it isn't that optimized for OpenGL. At least I've found it just as good for games FPS, multi-monitor, video acceleration etc as Windows
At least lm-sensors got an upgrade for Radeon 5XXX series, because the edge reading is for that series (and newer I guess), because they have a ton of sensors and the one on the edge decides. Also k10temp got Tdie and Tctl.

amdgpu-pci-0900
Adapter: PCI adapter
vddgfx:       +0.90 V  
fan1:        1004 RPM  (min =    0 RPM, max = 4500 RPM)
edge:         +44.0°C  (crit = +94.0°C, hyst = -273.1°C)
power1:       31.13 W  (cap = 120.00 W)

k10temp-pci-00c3
Adapter: PCI adapter
Tdie:         +60.4°C  (high = +70.0°C)
Tctl:         +60.4°C

You get this good result on your NVMe? I get the same on both (CPU bound and MB bound)

Went through the process of booting Windows for the first time in months (Insider Program, Fast Track). Only 1 reboot (long) reboot. Wow. Ran Geekbench in CLI while system idle, so fair. Linux is still winning. Also in Blender Benchmark but didn't have time to run the new version on Windows, but surely is.

Linux

Windows

olejon on 12 Jun 2020

Having sensors, especially their source code, is driving me to Radeon.
SAPPHIRE PULSE RX 5600 XT 6G appears among the best.
I've not decided yet how to partition the Sabrent nvme disk. Will I need Windows ? Especially to dig into the Ryzen tools and compare the _CoreFreq_ results with. Don't have an idea how to _clone_ correctly from previous SSD, the W10 part and it's boot sector.
I'm using Windows less than 10 times in a year. This OS became time consuming. Booting it once a while from a USB thumb key will be enough for me.
If such boot is feasible ?
I'll post back some Sabrent benchmark using the hdparm speed test...

cyring on 12 Jun 2020

Well I have Windows on a separate disk that was in my monster laptop (Windows accepted the total HW change). The only "annoying" thing is that I can't change it to a UEFI boot partition scheme, hence being able to set the BIOS to only allow detecting and booting UEFI bootable devices. Must now also have "Legacy OPROM".

Though, since Windows 10 is activated through your Microsoft Account now (make sure it is, if not convert it to it), I should be able to totally reinstall Windows 10 by downloading the Insider Program bootable ISO and install that and reactivate Windows without problems.

The same goes for you. You can always opt-out of receiving Insider Program updates, or choose the Slow Track, although the Fast Track has only been a benefit for me. No bugs, no instability. Just updates more frequently, something you want on newer HW.

Just register here with your Microsoft Account and you'll get a bootable ISO:

https://insider.windows.com/en-us/

Would definitely install Windows first (with Secure Boot DISABLED at install), then Linux. A user friendly distro should tap into UEFI just fine (or follow guides).

Since you use Windows so little, can't you just backup whatever is in you C:\Users folder (from say Linux, no cloning needee) and put it back again? Reinstalling programs shouldn't be a big deal?

IDK hdparm can't measure my NVMEs at all. GNOME Disks can. But not write speed when mounted, and not from Live USB either as the USB doesn't write fast enough. And Windows tools (for God's sake use ATTO Benchmark, only reliable) can only check if there's an NTFS partition on them.

Can always make some free space at the end, but well, Windows doesn't live on it so why... I trust GNOME Disks. It shows specification speeds, a little higher often (after BIOS + firmware updates, latter only possible from Windows...).

As for GPU choice can't help you more than buy one with a return policy that allows testing it as long as original packaging is not broken, test it and return if not satisfied. That's what I would've done.

olejon on 12 Jun 2020

BTW the EFI partition should be the first partition (doesn't need to be but MUST be witin the first X GB of the disk so stupid not to have it as first).

On Linux and set to be mounted as /boot/efi. I just set to to 1 GB as per some guides, but Linux uses only 6 MB on it. Better to just let Windows set it up. Uses FAT32 with boot, esp boot flags:

# parted /dev/nvme1n1 print
Model: Force MP600 (nvme)
Disk /dev/nvme1n1: 1000GB
Sector size (logical/physical): 512B/512B
Partition Table: gpt
Disk Flags: 

Number  Start   End     Size    File system  Name  Flags
 1      1049kB  1075MB  1074MB  fat32              boot, esp
 2      1075MB  1000GB  999GB   ext4

It's a hazzle to set up afterwards, which I did. Had to use some fancy fdisk stuff to fix the ext4 partition. Windows doesn't let you add free space at beginning, but GParted does allow shrinking the ext4 from there and it worked just fine. Just on 1 of my PCs did I have to do some magic from Live USB to make it boot again. Think it got messed up when setting the specific EFI GPT FAT32 partition table to to loop (partition table inside partition table) which is what Ubuntu does at least, using loop. Only certain tools show this.

# parted /dev/nvme1n1p1 print                                  Model: NVMe Device (nvme)
Disk /dev/nvme1n1p1: 1074MB
Sector size (logical/physical): 512B/512B
Partition Table: loop
Disk Flags: 

Number  Start  End     Size    File system  Flags
 1      0,00B  1074MB  1074MB  fat32

olejon on 12 Jun 2020

Some edits...

olejon on 12 Jun 2020

Thanks for all your tips
I have found this tutorial, in French unfortunately, to use the Windows Media Creation Tool to an 8GB USB bootable flash storage.

I will give a chance to a multiboot EFI/GPT partition.
I'm back in a while.
Cheers

cyring on 13 Jun 2020

It's easy. You go here and get the Insider Preview.

If you're not already an insider, become one, it's free: https://insider.windows.com/en-us/
https://www.microsoft.com/en-us/software-download/windowsinsiderpreviewadvanced?wa=wsignin1.0
Sign in clicking the upper right User Icon
Scroll down to Select edition
Choose First in List
Confirm
Choose language
You get links for 64-bit and 32-bit
FAQ with links to Flashing: https://www.microsoft.com/en-us/software-download/faq

olejon on 14 Jun 2020

Read above first.

Edit: Maybe installing Ubuntu 20.04 after Windows, which should set up everything correctly, then look at those partitions and mount points, and use those partitions for Arch Linux :-) which I suppose you still use.

Just DON'T delete Windows EFI stuff on the EFI partition(s) or modify them. Though I assume Arch Wiki has a good page on this, NOT needing Ubuntu install.

olejon on 14 Jun 2020

And again, read above-above and above first.

Stuff is under /boot/efi/EFI/ as standard. Can see the ubuntu folder for instance. Fedora makes a fedora folder...

On Ubuntu it installs various grub-efi-x packages. Key is the GRUB EFI packages, and NOT install the grub-pc-x packages.

This is Ubuntu default anyways...

Just this package below pulls in the necessary, if wanna switch, dependencies conflict with legacy GRUB so fixes that, as the dependencies pulls in all, like efibootmgr (indirectly) and mokutil, sbsigntool directly and all the GRUB EFI packages, and well why not the recommended secureboot-db.

I had made the mistake of installing the legacy GRUB (as part on minimalistic install and desktop from there).

I just uninstalled all GRUB packages and made sure I had the /boot/efi partitioned correctly and mounted, installed the EFI ones before reboot. APT correctly warns it'll make the system unbootable when uninstalling GRUB. For BIOS and GRUB setup (at least with Ubuntu's script, no path necessary then for installing or reinstalling GRUB) it's way better to have it UEFI compatible.

Note that MB manufacturers have said they'll probably get rid of the CSM legacy boot loader soon anyways (probably in your BIOS under Boot).

# apt show shim-signed

Package: shim-signed
Version: 1.37~18.04.3+15+1533136590.3beb971-0ubuntu1
Built-Using: shim (= 15+1533136590.3beb971-0ubuntu1)
Priority: optional
Section: utils
Source: shim-signed (1.37~18.04.3)
Origin: Ubuntu
Maintainer: Steve Langasek <[email protected]>
Bugs: https://bugs.launchpad.net/ubuntu/+filebug
Installed-Size: 1 388 kB
Depends: debconf (>= 0.5) | debconf-2.0, shim (= 15+1533136590.3beb971-0ubuntu1), grub-efi-amd64-signed, grub2-common (>= 2.02-2ubuntu8.1), mokutil, sbsigntool
Recommends: secureboot-db
Breaks: grub-efi-amd64-signed (<< 1.93.7)
Task: ubuntu-core
Supported: 5y
Download-Size: 344 kB
APT-Manual-Installed: yes
APT-Sources: http://archive.ubuntu.com/ubuntu bionic-updates/main amd64 PackagesDescription: Secure Boot chain-loading bootloader (Microsoft-signed binary)
 This package provides a minimalist boot loader which allows verifying
 signatures of other UEFI binaries against either the Secure Boot DB/DBX or
 against a built-in signature database.  Its purpose is to allow a small,
 infrequently-changing binary to be signed by the UEFI CA, while allowing
 an OS distributor to revision their main bootloader independently of the CA.
 .
 This package contains the version of the bootloader binary signed by the
 Microsoft UEFI CA.

Sub-package of GRUB is signed with Canonical's key which AFAIK almost all distros use, as Microsoft gave it to them.

NOTE: Made for Secure Boot ON yes, but it DOESN'T have to be on!

So why not just use it if wanna play with Secure Boot and CoreFreq to add to Wiki?

With everything in-kernel it can be on just fine, but say NVIDIA's driver must be signed, BUT:

Latest Ubuntu 20.04 let's you install NVIDIA's driver easy on install. Before Ubuntu had a solution, asks for password on GUI install, just choose 12345678, reboots into UEFI and sign with that, reboot and done. Should also work with the CoreFreq module signing that way if in correct directory (dkms...)

olejon on 14 Jun 2020

Thanks for all these tips.
I have a Windows 7 professional license upgraded to Windows 10 Pro. Installation is done onto a GPT 1GB EFI plus 192GB Windows data.
Now, using my _CoreFreq_ live image which embeddes the Arch Linux setup, I will follow directions of systemd-boot wiki

cyring on 14 Jun 2020

After playing a little with Ubuntu 20.04 using the Live USB, seems like this Ars Technica article is right on: Ubuntu 20.04: Welcome to the future

Just switching to the dark theme... Much nicer than macOS, at least when macOS doesn't run on a very PPI monitor. Seemed to work fine in programs like LibreOffice, VLC, AudaCity etc, you know programs using a good mix of toolkits. On the fly.

All the theming changes, all the collaboration + contributions with GNOME and Debian devs. So it's no more Ubuntu-stuff on top of things. More Ubuntu things adopted by GNOME. Already seen in Fedora but... Like the Control Center and Nautilus. Remember when it looked like macOS. Still trying too touch friendly. It still is but just less need for gnome-tweaks (none for regular users, just like a macOS user would need TinkerTool).

Maybe it'll be my first time using the default Ubuntu desktop. In fact it probably will. Wallpaper and terminal theming change. About it. With all the theming. Heck even as planned but also as requested by HP, Dell and Lenovo Ubuntu now shows its splash screen loader during the POST logo, just like Windows does.

That is, if installing on a laptop sold with Ubuntu as an option. Probably can be activated for any PC, but my workstation boots so fast I can't see anything before suddenly it's time for login :-)

All of this theming is FOSS of course. Only reason to mess around with it becomes to "stand out and not have a little different desktop than others". For GNOME Shell there's only 2 extensions as I could see. Dock and legacy tray icons (IIRC deactivated by default). Hiding dock and you got Vanilla GNOME basically... Which is the metapackage I use.

They finally having an agreement with NVIDIA to install that driver at installation of the OS is nice for people with an NVIDIA GPU.

Hotswap of the kernel seems even easier. Now Ubuntu requires a Ubuntu SSO account for this. But you know it's possible, and has been, for a long time, without. It just makes it so dead easy and stable that first time users are encouraged to use it. People used to Windows must be like "WTH do I need to reboot this system, really?" :wink:

And Software Center (GNOME Software but any compatible can use the underlying as you probably know). Well I prefer to install everything through APT. Maybe a stupid obsession as snaps are way more secure (Flatpak also supported, which comes with Fedora and snapd one click install). Even more fine grained permissions model than say Android.

The little I've seen makes it dead easy to get your program to the Software Center. If rejected, help to fix it. Usually devs adding access to files/folders the program doesn't need. Like the entire ~/.config when only needing ~./config/<program>. Even a lawyer chimes in if in doubt about naming/copyright.

Like the entire Microsoft Visual Studio for Linux is one click away. So is Spotify, Steam, Skype...

Just need more OEMs selling laptops with Ubuntu. This might be a first step.

BTW yeah the "What's New" celebrates the spin-offs too like Kubuntu, Ubuntu MATE etc.

Not that Arch Linux (and many others) isn't great, just won't see that on HP/Dell/Lenovos :P Like PopOS is already sold preinstalled, just not the big ones, yet.

Interesting.

According to this it works "A-OK" (I take it as it works A = Perfect) on Ubuntu 20.04 too (guide made for 18.04). Stack Overflow question says so.

https://blobfolio.com/2018/06/replace-grub2-with-systemd-boot-on-ubuntu-18-04/

Maybe I'll try it... For what? Faster boot? I never boot and if so never see GRUB anyways. As usual probably wait until the first point release, or even the second which comes with the HWE stack by default (that is when 20.10 is released + 3 months).

Login says:

Your Hardware Enablement Stack (HWE) is supported until April 2023.

Ubuntu 20.04 increases this to 2025 of course.

BTW for Windows, IDK about your motherboard's drivers listed on the website, but DON'T install the AMD chipset one that includes the graphics driver, unless that's the only chipset package there is. ASUS has one without. Just messes up things when you install the latest from AMD.com, where you should go for the "Optional" (latest) Adrenalin driver package, if you have an AMD GPU (if not probably best to let Windows 10 install them, that is if it actually has them all in Windows Update, but on Fast Track I'm pretty sure).

And always choose Advanced/Extra options during install and choose a "Hard install" (something like that). Completely uninstalls the previous. Not doing that I've seen an error various times when finished. Although works afterwards, really hard to see if it got installed if not comparing Windows Device Manager version before and after. The Radeon Software is pretty great. It can check for updates for you, just make sure to set it to check for both Standard + Optional.

olejon on 14 Jun 2020

Would be really cool to see your 3950X results from Blender Benchmark (CPU). For comparison. Stresses my system more than Conic Compute.

Linux wins with a whopping 1 minute and 3 seconds on classroom. Seems to be ~9.3% win on average.

Just a launcher you can place anywhere. Data it downloads is simply placed in ~/.cache/ as it should... At first launch it downloads the necessary parts of Blender (use latest suggested version), and well, I just use the bmw + classroom renders.

Download: https://opendata.blender.org/

Linux

Windows

olejon on 14 Jun 2020

I will send you blender results when setting Arch is up.
So far, I have reinstall Win10Pro to "force" a GPT partitions table. I guess you warned me about this.

So far, I have installed systemd-boot with one Arch entry. Choice of UEFI entries is off course made straight from the fancy ASUS firmware. I plan to add one entry per kernel I'm working with.

Next, I have to set up all the development lab...

Meanwhile I have played with Ryzen Master which is a powerful software but its CPU overhead is about 200 MHz, whatever the Windows performance policy is (balanced, high, saving, policies made by Microsoft or AMD)
Booting the _CoreFreq_ live shows a CPU overhead less than 10 MHz, but also less watts, voltage, lower Core temperature.
Interesting results.

P-States are working, no crash or locks, no TSC issues. Only some compute issues due to the fact that I don't have the registers to recall the original P-States values, once changed.

For the next feature, I wish to measure the c-states, if general counters can be programmed for this; and specifications available.

cyring on 14 Jun 2020

👍1

Btw, this issue being solved, feel free to close it, but keep it private for our tchat.
Regards
Cyril

cyring on 14 Jun 2020

Yup. Same here. UEFI shows simply Ubuntu (Corsair + Model).

Since my Windows lives on a separate SATA SSD, I don't bother to convert it from Legacy. Must just set UEFI to, for internal disks, detect "UEFI + Legacy OPROM", and press F8 when booting and choose that disk.

Little annoyed at first when update-grub, which is run when new a new kernel version is installed detected Windows as well... My temporary solution of unplugging the SATA cable wasn't sustainable :P

Fixed setting in /etc/default/grub:

GRUB_DISABLE_OS_PROBER=true

So it doesn't check for other OS'

olejon on 14 Jun 2020

Hardware done

Intel, left - AMD, right

cyring on 14 Jun 2020

🎉1

_CoreFreq_ build ready

As Ryzen Master says I have stressed the best Cores which are CPU number 0 and its SMT 16

Indeed best Cores can reach 4.7 GHz, not the other CCX

cyring on 14 Jun 2020

I saw your Blender score, pushed the button; and said to myself: let's have a coffee
What a shock when I came back !
blender_bench

Recommend you to try again under Linux...

cyring on 14 Jun 2020

LOL. But I did post Linux & Windows. You thought it would take long? If your link sucks then the download would probably take longer :fr:

I suspected it would take approx. 3 times less, although not 3 times the cores. Almost right.
What Frequency & Temperature do you get with Conic Compute, first choice in list?

Assume you have latest BIOS
Have you enabled PBO? It can be done in UEFI (but you must know what settings, then it'll show in Ryzen Master as OC Mode = Auto Overclocking) or just use Ryzen Master, setting whatever profile to Auto Overclocking then Apply & Test, which will set it to 100 MHz (max is 200 MHz but makes no difference, no tests I've seen see benefit over 100 MHz)

For me it enables in UEFI under "AMD Overclocking" these:
- Memory overclocking: Enabled (Rest "Auto")
- Precision Boost Override: Enabled (1 drop-down set to 100 MHz, rest "Auto")
- That's it, does not change the PBO settings which are on the main AI Tweaker section of UEFI (seems like a setting that does nothing)

olejon on 14 Jun 2020

LOL. But I did post Linux & Windows. You thought it would take long? If your link sucks then the download would probably take longer

I suspected it would take approx. 3 times less, although not 3 times the cores. Almost right.

Got familiar to wait with my Xeon
Now between the FTTH and Ryzen, I can't barely have a nap

What _Frequency & Temperature_ do you get with Conic Compute, first choice in list?

CoreFreq_Ellipsoid

Apparently the Celsius 36 is doing a great cooling

Assume you have latest BIOS

Have you enabled PBO? It can be done in UEFI (but you must know what settings, then it'll show in Ryzen Master as OC Mode = Auto Overclocking) or just use Ryzen Master, setting whatever profile to Auto Overclocking then Apply & Test, which will set it to 100 MHz (max is 200 MHz but makes no difference, no tests I've seen see benefit over 100 MHz)

For me it enables in UEFI under "AMD Overclocking" these:

Memory overclocking: Enabled (Rest "Auto")

Precision Boost Override: Enabled (1 drop-down set to 100 MHz, rest "Auto")

That's it, does _not_ change the PBO settings which are on the main AI Tweaker section of UEFI (seems like a setting that does _nothing_)

Beside DRAM set to 3600 MHz and CL 16, all BIOS options are in AUTO mode.
RM confirms the OC mode. But I still have a lot to compare. Between the BIOS bits and the _CoreFreq_ readings... ie. setting BIOS options to [Enable] | [Disable] rather than [Auto] then query values in driver...

Here is one of the strange things I have noticed so far:

CC6 [ON]
CC6 [OFF]

Almost 1 MHz lost when disabling CC6

cyring on 15 Jun 2020

CC6 _behavior_ confirmed by the power consumed; voltage to observe.
2020-06-15-005002_484x550_scrot
2020-06-15-005033_484x550_scrot

EDIT: Stressing a single core, there's a delta of 15W on the whole package.

cyring on 15 Jun 2020

Question: According to Ryzen Master's documentation:

Gray and Gold stars are fastest per CCX and Gray Dots are second fastest per CCX

Here are screenshots from Ryzen Master and CoreFreq Topology. You can see CCX0 has 1 Gold Star and 1 Gray Dot, while CCX1 has 1 Gray Dot and 1 Gray Star.

Which in the list should perform best in a CoreFreq: Turbo Select CPU?

I'm pretty sure CPU 000 is the one with the Gold Star as it performs best. Then I get a little confused...

Ryzen Master

CoreFreq

olejon on 15 Jun 2020

Be aware that _CoreFreq_ numbering is zero based.
Yes, you will run the Turbo Select CPU function to stress the CPU number zero in _CoreFreq_ which is the best Core according to RM
One thing I don't get is how RM find this best Core ?
At first, I would imagine RM stresses and measures each Core during its startup; but I don't notice such action on Processor while launching this software.
My guess is that some registers are storing those factory informations.
Found nothing in specs.

cyring on 15 Jun 2020

AnandTech: AMD Clarifies "Best Cores" vs "Preferred Cores" Discrepancies For Ryzen CPUs
Think I got it. What RM shows is right according to MHz achieved in CoreFreq. I know it's 0-based of course. Just don't get the CCX/ID columns
So say stressing 000 gives same result as stressing 006 (same core)
RM can show when a core is sleeping. IDK if it actually is with its Windows drivers or it's just so low on MHz it says "Sleeping"
BTW I was stressing the CPU using CPU-Z at the time of screenshot. Had nothing to do with RM!

Oversized package for the 3950X, basically as those with a stock fan, but MAN that Noctua package size! :fearful:

olejon on 17 Jun 2020

Congrats. Time for hardware build, now.
About SMT Cores: AMD should have the micro-architecture answer. But you will have to run different load & idle cases and observe resulting Core states and sensors.
With Nehalem, stressing one physical Core has "effects" on its SMT sibling: like voltage, temperature.
Zen sleep states are my next duty...

cyring on 17 Jun 2020

👍1

Zen sleep states are my next duty

Nice :+1:

Yeah I guess, no rush. I hate changing the CPU (and from now on MBs without an I/O block, really really hate the I/O bendy shields I sometimes forget to put in place until to late...).

If it wasn't for coolers it would be 10 seconds but... When my GF can help. With my chassis' backplate for the CPU (chassis has 2 doors) it is much easier with 4 hands, especially 4 eyes to see the screws come through correctly. Want it perfect the first time, after applying thermal paste and all.

I remember I did consider buying your MB, for the Wi-Fi and additional 2.5G LAN, but wasn't worth double the price. How's the Realtek working? I bought early on and remember reading some on reddit saying it worked but not exactly perfect (yet). Could crash or whatever.

Also the whole ROG, searched it to "Republic of Gamers", and "HERO" RGB put me off :laughing: instead of my subtle one, the white I/O block on mine makes it kind of look integrated into the chassis too.

Later we can compare benchs, and see who got luckiest with the silicon :wink:

olejon on 18 Jun 2020

I have made my components choices with the help of those awesome lists:

https://docs.google.com/spreadsheets/d/1wmsTYK9Z3-jUX5LGRoFnsZYZiW1pfiDZnKCjaXyzd1o/edit?usp=drivesdk

https://docs.google.com/spreadsheets/d/1c3RTF_ZPjep-Zfimgoca2Ef1gSjZM0rSHVWLkknbfUI/edit?usp=drivesdk

https://docs.google.com/spreadsheets/d/1dsu9K1Nt_7apHBdiy0MWVPcYjf6nOlr9CtkkfN78tSo/edit?usp=drivesdk

https://docs.google.com/spreadsheets/d/18aMZfSQFkVUQWjVt6RSFDh866D1wZlWAxSF7mAHWkbA/edit?usp=drivesdk

https://docs.google.com/spreadsheets/d/1Smj5dh97n32wJqm5dkdDcQt8ID7vH52-lKzaaXUUQx8/edit?usp=drivesdk

At first, I was looking for Pro components, like an EPYC, Threadripper, second; but Home-server components , for a Workstation form factor, was not so available. Cooling is also a matter of worrying. Thus gaming components, but with the best vrm motherboard, capable for the highest cores count among mainstream processor.

So far I'm not considering the ROG devices: network, wifi, BT, sound are just doing their job. Ryzen is my focus. ASUS BIOS has features I can compare _CoreFreq_ with.
This PC has to last a long time. Previous ROG Gene II, upgraded with W3690, is still alive 10 years after. I thrust this brand. Only 3rd MB was RMA and fixed by Mfr.

Diving into Zen datasheets again, I don't find specifications of idle states I can program the PMC counters with.

cyring on 18 Jun 2020

Oh, that's indeed some nice Docs.

Yeah the for the vrm cooling on your MB is probably better. Took a look here and seems pretty good/similar. Only benchs will show.

Any big difference? Seems to be the top ones under heatsink and the side ones on both next to the side heatsink (and some next to RAM?)

Images comparison.

ASUS PRIME X570-PRO

ASUS ROG Crosshair VIII Hero

According to This doc - found on Reddit but one of yours

My Prime X570-Pro ticks all the Green Boxes. But not sure how to interpret under Vcore VRM:
- Prime X570-Pro = Phases 4 - 3x SiC639 50A Power Stage
- ROG Crosshair VIII Hero [Wi-Fi] = Phases 7 - 2x IR3555 60A Power Stage

Just saw in Reddit (which has the Docs link), one reply to using the 3950X on a cheap MB that is no longer sold - here at least.

If he doesn't manually OC (stupid thing anyway with these CPU's) the worst that will happen under heavy load is the CPU won't clock as high due to insufficient VRM power (cooling). At least that's the way I understand these processors work.

olejon on 18 Jun 2020

Nice when everything works the first time and POST says "New CPU Installed, press F1 for Setup"
As always with new heatsinks, difficult to fasten the first time, had to take it off and lay the chassis down, hence fearing the Noctua thermal paste would need cleaning, but I had put it on perfect so it covers the surfaces perfectly, not too much, not too little
Often the packages are way bigger than the content, like the 3950X is the size of a big stamp in a standard box for coolers, but the Noctua heatsink - wow - I can't see my nice subtle I/O block's RGB anymore, looks like there's an engine in there now... but RAM and PCH fan lights, yes (just color pulse)...
And I haven't put on the second fan, with just the one in the center, the system struggles to reach even 70 Celcius (have to use Blender Benchmark and can probably avoid it by ramping up the CPU fan to 100% at a lower temperature), also the ambient temperature, now it's 30 Celcius outside and inside, so... Just the one fan works excellent, and so much quieter than the stock for the 3600X, and the idle sound, as it should be for that size, could ramp it up to 40% duty cycle at idle, no difference, and not using the noise-lowering cables that come with the package...
I think the second fan will crash with my RAM chips, any RAM chips on an X570 ATX MB, unless you mount it vertically (really just 2 mount options because apparantly there are 2 types of backplates) which is possible, but then air will just be dragged in from nothing (top) and down to the GPU...

This is the best comparison IMO, against an Intel i9 XE with more Cores (18C/36T)
It also beats the AMD Threadripper 1950X 16C/32T

olejon on 19 Jun 2020

Definitely it looks like a car engine. Can't way to see your Conic temperature.
Remember, I got 70C with the Celsius S36, not at its highest fans rotation.

My HTPC is using a low profile Noctua nh-c12p to cool the i7-920 processor. Yesterday, removing dust, I changed TIM with MX-4 which resulted to temperature as low as with Noctua TIM.
Off course I'm using a TIM solution cleaner for both, processor and heat-sink, before applying a "big" drop of TIM, but it becomes difficult to get a thermal benefit among the TIM top 10

cyring on 20 Jun 2020

Lol, and definitely is a good one too
Remember I just have the middle fan installed, but nice to have a spare one then. Doubt it would've made much difference considering temperatures below and previous experience with this chassis' airflow
IDK which Chassis + X570 MB is made so that one can use both without clashing with the RAM chips.
SSH screenshots from my phone:

Idle Temperature - This is just 4 degrees above Room Temperature!

Conic Temperature - Approx 2 minutes - It never goes above 70

Note: When running Blender Benchmark all cores stay at almost 4000-4100 MHz, seems like it gets more out of the CPU than Conic despite stressing more, see next
Blender Benchmark manages to get it to some 71-72 degrees, and cores run at higher MHz than seen in the screenshots above, so as before Blender Benchmark is the best stresstest for temperature and MHz
This Noctua 140mm fan is so quiet and nice I can set it to a higher duty cycle when idle and also to gear up 5 degrees before (65) the chassis fans do at 70 degrees, so it's very silent running Conic
Ryzen Master shows % of 142 Watts, but with the 3950X CoreFreq actually shows a significantly higher Package Watts than the TDP of 105 Watts, when stressed (logical?)
This wasn't the case with my 3600X with a TDP of 95 Watts, where Ryzen Master showed % of 128 Watts, and CoreFreq Package Watts stayed at around 87-90 Watts when stressed (seemed logical)
And sorry it wasn't an Xeon I know... But it's the fastest Intel one in CPU-Z's reference list by far, and it's 18C/36T so...

More benchmarks and comparisons here:

https://www.olejon.net/files/X570-3950X/

olejon on 20 Jun 2020

Around 70C is a good result, especially if fan stays quiet
Can you reach 4.7 GHz single Core ?
You Blender result is 10 seconds more than mine in classroom (worst w/ Windows): do you have CPB disabled in BIOS ? or DRAM speed and latency matter ...
Nice Linux Software of the day

OpenRGB works fine with Aura

cyring on 20 Jun 2020

👍1

Yes I get a little above 4.7 GHz on "gold star cores"
Can't remember having touched CPB, will check tomorrow
Yeah, your faster RAM came to mind, it definitely has an impact and with 3600 MHz one can achieve even better performance not using the default Infinity 50/50 (1800/1800 MHz), but also the newer kernel you run came to mind... Can check with a live USB easily later if there's gains there, if so Ubuntu 20.04 upgrade probably sooner, BUT not that it should be noticable outside benchmarks ehh :wink: Got a graphics stack update recently and made the radeon driver even slightly better, small things also like I don't have a splash screen of course since system boots so fast, and a tiny flicker when GDM is loaded just disappeared with that update, no kernel, just X.org, drivers and Wayland...
Nice! Will check but for now my set settings in Windows are fine. But it should exists and be an open standard. Kind of thought it was. Many use it. The ASUS software for Windows can tap into Corsair's RAM chips' RGB settings for instance
As I bought early, and default was so nice, didn't bother to search, but knew it either existed already but probably buggy, but knew it would come... Like:
Was pretty surprised when I found this for Razer peripherals, just driver and daemon and let others make nice GUIs or just CLI using it. Totally works but my keyboard only has backlit keys with one color (that's why I bought it, good keys, backlit and no RGB fancy), so can be used to change the level which of course can be done by Fn + Keys anyways - also the just annoying "Fade in/out" effect can be applied, and speed. Can look cool when idle but absolutely not when using it. But bet it's nice for other keyboards with more features. Razer's own Synapse software is worse IMO (and the remapping of keys, why, better let the system do it), especially on macOS when I used my MacBook as a workstation temporarily living in Spain

olejon on 20 Jun 2020

Ryzen Master shows % of 142 Watts, but with the 3950X CoreFreq actually shows a significantly _higher Package Watts than the TDP_ of 105 Watts, _when stressed_ (logical?)

This wasn't the case with my 3600X with a TDP of 95 Watts, where Ryzen Master showed % of 128 Watts, and CoreFreq Package Watts stayed at around 87-90 Watts _when stressed_ (seemed logical)

Please check this AnandTech 3950X review
What it says:

At its peak, the 3950X draws 137 W for the cores when 10 cores are loaded. The chip as a whole hits ~144-145W at that level, well above the 105 W TDP rating on the box ...
Now _CoreFreq_ power estimation.
Stressing Processor with Conic > Hyperboloid of two sheets
Look at the whole Package measurement.
Same conclusion ?

cyring on 20 Jun 2020

Checked the CBP (Core Boost Performance I assume) and it was "Auto" == Enabled. Setting that to Disabled obviously made it slower
I already have DDR Gear Down == Disabled (remember that from YouTube video being something that could give make benchs/games perform *a little better, BUT not CoreFreq stressing`
Did get about ~135 Watts
Also set some DDR Power Down/Saving == Disabled, but see no advantage of that. Surely was the SoC Uncore in same UEFI section AMD Overclocking which makes the big difference here, just look :open_mouth: :

If you think this is "cheating", well, not with this Noctua fan. Reason I bought it instead of the slimmer also just for AM4 is because it has the logo in the image below for the 3950X, so my fan is listed as best turbo/overclocking headroom
Yes it gets hotter as you can see, and my 3600X didn't like this with the stock fan, obviously. In fact the PC suddenly rebooted a couple of times when stressed, but not always... :thinking:
Good news: Total chassis cooling seems to manage this just fine, not going above 86°C, and idle temperature isn't affected
Have to check Blender Benchmark later when actually using the desktop, not just SSH
As you probably know AMD considers 75°C as "high" as per sensors but that means little, Intel uses 85°C. To reach the "hysterical" point AMD has 95°C before throttling or shutting down, and Intel 100°C (Intel always gets hotter)

olejon on 20 Jun 2020

More than 195W is scarry and with 82°C, I wonder when thermal throatling is going to engage ?
I'm gonna review the driver to see if I had set a thermal sticky bit into code, which translates into HOT being highlighted.

cyring on 21 Jun 2020

First of all it's an official Enabled/Disabled setting in UEFI and the Help Text simply says something like "Can boost performance at the cost of power saving"
But when Idle, Temperature and Watts = normal
Also my 3600X with the stock cooler often got above 75°C, which AMD considers "High" for all these CPUs, and I've never had problems with the temperature being about what Intel (85°C) or AMD consider "High"
A MacBook Air, many other laptops and iMacs too with Intel reach close to 100°C ("Hysterical") quickly with just some moderate multitasking, as easily seen as they start to throttle. Then CoreFreq correctly for my MacBook Air with Haswell reports 3 indicators for HOT, but there's no way to avoid that and they're built that way, and BIOS set to throttle instead of shut down for CPU protection
When my workstation ran Skylake i5 I set all cores to Turbo and overclocked all so the CPU reached 95°C under stress without problems, just used a sensor tool to simply reboot the PC if I wasn't using it, and for some reason say some process went crazy and it reached 96°C (never happened but tested and worked fine)
I have only read that the "High" value is a temperature you generally want to stay below but it definitely can go higher in periods, just prolonged close to "Hysterical" can damage it over time, but reaching "Hysterical" causes the PC to shut off (or throttle if that's set in BIOS)
If any Intel laptop goes 85°C+ constantly, it can't be a problem for an AMD desktop CPU to go 82°C for a few minutes? Only benchmarks cause this
And regarding benchmarks, Blender Benchmark did NOT cause an even higher temperature BUT 203 Watts. Temperature actually floated at a slightly lower one than Conic IIRC
Now, Watts is another thing. Blender Benchmark showing 203 Watts, wow. My PSU is just 500 Watts so kind of surprised, so:
Must check Ryzen Master if CoreFreq's reading is right
Also must check if Ryzen Master gives any warning when running Blender Benchmark on Windows
BTW with this I got exactly 1 second less on both your results from Blender Benchmark :smiley:
Will read a bit more about this option before stressing the system again, but it's nowhere close to this at regular use, it is now at 14 Watts and 35°C as ambient is 29°C

olejon on 21 Jun 2020

Thanks for your temperature calibration effort. Can't wait to see your results.

Meanwhile I've checked code and driver has no ThermalTrip query in families 17h & 18h.
Thus no HOT will appear with Zen processors.
Perhaps kernel is logging it.
I will search for such thermal status register which may be queried onto SMU

cyring on 21 Jun 2020

Many bullet points. Please read them all.

You made me curious although I haven't slept yet :stuck_out_tongue:
First, correction, according to k10temp "HIGH" is 70°C, BUT forget that because:
Have to look at Ryzen Master, which only lists 1 temperature and that's called Limit 95°C, as said before. Goes for all the Gen 3 CPUs
I ran a full Blender Benchmark classroom and the system was totally stable
Ryzen Master did not even reach Orange color for temperature or any other
I am almost 100% sure I saw the Orange color on my previous 3600X when stressing that
In fact I manually overclocked my 3600X to Ratio 41 (4100 MHz), just 1 above standard 40, and when stressed it didn't like that, temperature rose and the system simply shut off, and I could without problems boot into UEFI and revert that setting to default. Same with my Skylake i5
Temperature towards the end was at 85°C, but still Yellow color and NO warning
As said, on Linux that temperature is lower during the same Blender Benchmark run
Ryzen Master now shows 395W, and 50% of that is 197.5W as in Screenshot 2
Ryzen Master Screenshots when Idle (almost, I did some things) and During Blender Benchmark (also see points below these):

Almost Idle

During Blender Benchmark

Absolutely no warnings from the kernel when pushed up above 80°C. I checked with dmesg, journalctl, sensors (which shows an alarm/alert for certain HW) and used grep on ALL files in /var/log and there was nothing matching any relevant keyword
The last message was simply that CoreFreq was loaded, others were standard stuff from server services and desktop environment
Blender Benchmark classroom finished faster on Windows as well, but result isn't valid as I had Ryzen Master open + I took a screenshot during the run
Blender Benchmark is very sensitive to any other stuff going on (system should be totally idle), and it stresses both Windows and Linux so much for me that the desktop lags a bit and is a bit slower to open say the file manager
The below says "... expense of idle power savings", which may be true for the RAM but I don't see the CPU being more active when idle (and there's no integrated graphics as you know so that doesn't count)
Screenshot of the Name and Description of the Setting in UEFI. I see nothing that screams Danger! Danger! Do you?

Both Red Rectangles say the same

olejon on 21 Jun 2020

Very interesting investigation you did, although I will go very slowly on those settings t'il we have a clear understanding on what is going on, and their side effects. I'm still not sure if hardware is well protected against over-voltage, over-consumption and so on.
Don't fry your brand new Ryzen !

This morning I went through the last PPR, this document revision applies to model 20h which concerns APU to my understanding.
What we found is a list of registers not unveiled in previous families. Most of them can be queried through wire protocols: SM-Bus, i2C, GPIO, etc
I bet AMD has some MSR to ease the readings but those are not public yet.

We find wanted registers a bit every-where, sometimes through :

ProcHot, ThermalTrip : on SMIstatus, SmiControl5 & Power Management (PM) Registers
C1e : PMx07E [CStateEn]
IdleCount (aka "idle count from the latest monitored period") at MISCx060 [IdleCntrl] (FCH::MISC::IdleCntrl). Is this what RM called Sleep from ?

Have a rest, you deserve it !

cyring on 21 Jun 2020

👍1

Don't fry your brand new Ryzen !

How can that happen when:
I ran a FULL Blender Benchmark with ALL the scenes, more stress regular usage will ever do
SSH + CoreFreq to monitor Temperature and Watts
Temperature: 78-85°C
Watts: 173-203
The scene victor took very long to load ("Warm up"), it seemed like it tried several times so I was on alert, but CoreFreq showed something was going on at least, and it did load and logical as it was the definitely longest one, but Watts during dropped to the lowest during that at times, to 173 Watts
In fact 203 Watts only seen during the first 2 ones: bmw27 and classroom, but never above 85°C during those
If going to 86-87°C it only took 1-2 seconds to drop back down to 85°C
During pavillion_barcelona - I think, if not koro - it managed to push it to 87-88°C, but not above. Watts was well under 200 then
BUT: If running those scenes again alone, it stays at 85°C and around 200 Watts, so I suspect it was because they load quickly and maybe didn't allow the CPU to cool down a little during meanwhile, maybe
So maybe I have to repeat this to see if it happens again during 1 of those 2...
Again my PSU is only 500 Watts, as I don't have an insane GPU, and don't plan to, as they freakin' now require a whopping typical minimum of 750 Watts PSU as 2 PCIe Power Cables are required (8 + 8 pins = 16 pins). Why didn't they fix this with PCIe4? Enough power in the slot please
Or GPU manufacturers, please look at what Apple fits into some square centimeters if not millimeters on iPad Pros with NO fans...
Really goes for all cabling, 1 standard please. I hope they look at USB-C + Thunderbolt, and use something like that for every cable (I like the design of SATA cables so e.g. they can keep that design). I've used the most fancy enterprise docks and they support tons of monitors, audio and I/O of all types, from 1 cable from a laptop

olejon on 22 Jun 2020

Read the above first
I can only find ASRock documenting "Uncore OC Mode" and it seems to have to do with RAM and Infinity Fabric only
I've not touched any of these manual settings, just turned off Gear Down which should be totally safe, and another Power Saving setting for RAM which is a sub-menu under that part again
Still system performs better. So is it simply the fact that it now actually uses my RAM to the max as D.C.O.P. = Enabled or what?
All people discussing this setting also do manual overclocking (entering values manually) but I don't

This does NOT seem to discuss that setting but rather manual overclocking in the AI Tweaker part of the UEFI, although maybe interesting for you? Remember to see the 2 screenshots below too.

"OC mode" is automatically engaged when P0 frequency is raised above base frequency. In OC mode, the SMU inside Ryzen CPU is disabled, which means sophisticated power management and power saving features are not functional. The presense of OC mode can be verified by calling GetCurrentOCMode() via the RyzenMaster SDK or code 0C via a Port 80 debugger card. HWiNFO has 3 voltage readings related to CPU. If it's individual core VID or motherboard Vcore that is scaling down to 0.2 Volts, then that's expected. What's important is "CPU Core Voltage (SVI2 TFN)". I think this value can tell whether the CPU has gone into "OC mode" or not. If this value is not scaling down, then it most likely means OC mode is engaged

Screenshots from 2 Manuals for ASRock MBs. Not much info:

PS: I've NOT touched this voltage!

olejon on 22 Jun 2020

Read the above first

I did but I don't feel comfortable with the power consumed.
173-203 are far too high to my taste
See my way in next post

I can only find ASRock documenting "Uncore OC Mode" and it seems to have to do with RAM and Infinity Fabric only

I'm not familiar with the Zen Uncore to tweak it yet. Especially its voltage limit.

I've not touched any of these manual settings, just turned off Gear Down which should be totally safe, and another Power Saving setting for RAM which is a sub-menu under that part again

I've also noticed this one and I don't know what's behind the hood. For now I have stabilized my DRAM at its OC factory mode and it will be easier to understand those extra settings when I will have a clear understanding of how the Ryzen Memory controller manages the [sub]-timings .
So far, I have no specs how to query the configured DDR speed, geometry, latency, and so on

Still system performs better. So is it simply the fact that it now actually uses my RAM to the max as D.C.O.P. = Enabled or what?

All people discussing this setting also do manual overclocking (entering values manually) but I don't

I have chosen D.C.O.P(XMP) to auto configure the G.Skill kit.

"OC mode" is automatically engaged when P0 frequency is raised above base frequency. In OC mode, the SMU inside Ryzen CPU is disabled, which means sophisticated power management and power saving features are not functional. The presense of OC mode can be verified by calling GetCurrentOCMode() via the RyzenMaster SDK or code 0C via a Port 80 debugger card. HWiNFO has 3 voltage readings related to CPU. If it's individual core VID or motherboard Vcore that is scaling down to 0.2 Volts, then that's expected. What's important is "CPU Core Voltage (SVI2 TFN)". I think this value can tell whether the CPU has gone into "OC mode" or not. If this value is not scaling down, then it most likely means OC mode is engaged

Screenshots from 2 Manuals for ASRock MBs. Not much info:

PS: I've NOT touched this voltage!

Those Windows tools are so lucky to get so many specs details from Manufacturers
I'm trying to stick to Linux and to offer an Open Source software. Road is long.

cyring on 22 Jun 2020

Correction: It is the scene fishy_cat that causes the highest temperature. See if you can replicate that, no matter settings
I ran the full Blender Benchmark with Ryzen Master open and saw the exact same behavior as on Linux with CoreFreq open, all of Temperature / Watts / MHz
Sat there during the whole run on Windows ready to close Blender Benchmark in case Ryzen Master showed 89°C, BUT it never did, as in Linux, reached 87-88°C during fishy_cat, other than that at 80-85°C
Speed gain: From a little above 3800 MHz to a little above 4000 MHz, not massive but clearly better results
Totally stable on both Windows and Linux and no alarms/alerts
I also have enabled D.C.O.P. and not touched any of the values it sets (UEFI shows what you have changed when you choose "Save & Reset")
I found my "Manual OC"!
I've always wondered why there are 2 settings for Precision Boost Override (PBO)
Ryzen Master only activates the one under AMD Overclocking
The 1 in AI Tweaker > Precision Boost Override is what makes the difference
Using the Drop-Downs to change both from Auto to Enabled + 100 MHz, the same as in the other, and BTW Ryzen Master still shows PBO is 100 despite setting it both places
Again, I have not entered any value manually, just Drop-Downs
Take a look at the differences in PPT, TDC and EDC, from 142 W to 395 W, NOT that it uses that, it uses 49-50% of that:

With both set to `Auto`

Set to `Enabled` and `100MHz`

This is UEFI setting alone causes that big difference in RM's Max Values, not that they are reached at all

In fact, if both set to Auto as in Screenshot 1, a Blender Benchmark causes PPT, TDC and EDC to go to Orange Color, almost 100%
Again, not that it does anything, Temperature is the SAME, system is stable and it does not show any sign of wanting to shut down, no alarm/alert, nothing
I think I will keep this "Overclock" on, as it's virtually impossible to reach above 70-75°C during regular use. Only benchmarks cause this... :man_shrugging:
If it goes wrong for some reason, I will put back in my 3600X and let you know :smile:
As Ryzen Master does NOT drop the MHz lower than when using CoreFreq, I guess a comparison is fine. Linux beats Windows, no surprise, especially in victor as it is the longest running one, but also the one with Lowest Temperature, actually:

Linux

Windows

olejon on 22 Jun 2020

Power tuning

BIOS settings: all `[AUTO]` beside a few

D.C.O.P standard with XMP but Memory Frequency set to 3600MHz
Applying a voltage offset decrease to use the minimum juice when CPU is idle but 4.7GHz single Core still capable
As automatically set by XMP, DRAM voltage is kept at 1.35V
Using Conic Compute... > Hyperboloid of two sheets which is the most _computing_ algorithm, I'm getting ~ 145W Package with a max Temp of 67°C at 1.3V
BIOS: PBO disabled, XFR disabled, CPB enabled

200622120949

SMT available
DDR timings: Beside Manufacturer's 16-16-16-36 values, all set to [AUTO]
CPU Power Thermal Control set by first POST. I don't know what's the purpose of this setting ?

Day to day working state

Single Core can go as high as 4.7GHz at 1.5V
An idle session such as Programming, Web surfing, PDF readings, etc claims less than 1V

cyring on 22 Jun 2020

That came quick! Remember to see the one I just commented.

Then we only have 2 differences in UEFI:

I have NOT touched CPU Core Voltage Offset. No manually entered value at all
I have PBO enabled to 100 MHz both places
Ryzen Master shows no PBO difference because of that, says Boost Override CPU = 100 and OC Mode = Auto Overclocking, which sets the UEFI AMD Overclocking setting automatically, but not the one in AI Tweaker, which for you is called Extreme Tweaker (same settings though)

I see you only get 1 reading under CoreFreq > Voltage, while I get for all of them
You know why probably?
You get a slightly better speed for Hyperboloid of two sheets, while I get a slightly better for Single Core, and both are totally stable in MHz
I see virtually no fluctuations during the full Blender Benchmark, all cores stay at exactly the same frequency, around 4065 MHz (think it varies a little depending on scene)
Screenshots of the same:

Temperature stays at 50°C

Temperature eventually reaches 80°C

Idle less than `1V` - 33°C with Ambient 27°C

olejon on 22 Jun 2020

How _CoreFreq_ is reading a Vcore of 0.2V : is it a voltage you set in BIOS ?

cyring on 22 Jun 2020

You mean the one at bottom right? Saying V[0.20]. Is that bad or anything?

I'm not into Voltage tuning AT ALL.

I haven't set ANY integer or string manually, only Drop-Downs. Only the D.C.O.P. changes automatically what you already know:

RAM voltage to 1.35
RAM timings, for me 16-18-18-36
RAM frequency for me 3200 MHz

Yours show V[0.91]. What's the diff?

olejon on 22 Jun 2020

You mean the one at bottom right? Saying V[0.20]. Is that bad or anything?

I'm not into Voltage tuning AT ALL.

I haven't set ANY integer or string manually, only Drop-Downs. Only the D.C.O.P. changes automatically what you already know:

RAM voltage to 1.35

RAM timings, for me 16-18-18-36

RAM frequency for me 3200 MHz

Yours show V[0.91]. What's the diff?

OK, so it's a _CoreFreq_ bug
Adding it to the todo list

cyring on 22 Jun 2020

Which reading? Yours or mine?
Could it be because I have Uncore OC Mode enabled?
A bit tired of rebooting and checking, people also use my server services during the day

olejon on 22 Jun 2020

Hmm, look at previous IDLE Screenshot:

https://github.com/cyring/CoreFreq/issues/186#issuecomment-646917760

It says V[0.91]
Suggesting it's because of Uncore OC Mode + AI Tweaker PBO which wasn't on when that screenshot was taken

olejon on 22 Jun 2020

Manual OC

Let's do things differently

BIOS: we're disabling CPB, PBO, XFR

Vcore is restored to [AUTO] which means that Processor posts at less than 1V
200622151135
200622151149

Ready to reboot

Doing the OC manually

Idle state

CoreFreq_Manual_Idle

BOOST register is confirmed OFF
_To answer to your previous question: Voltage Core scope is changed to Package in Settings menu_

Experimental mode

Let's increase the P-State ratio to 42 for all CPUs and stress my best Core
Let's go straight to AMD specs with a boosted ratio of 47
But can we go above ?
Let's try one bin

This 4.8GHz frequency appears very shortly but fallback immediately to 4.7GHz
Like something is monitoring against this P-State changed.
Interesting isn't it !

cyring on 22 Jun 2020

I suppose :man_shrugging: :smile:

You're using the CoreFreq CPU driver, no?
You have CPB 2 places in your UEFI? Maybe I do too but haven't looked, just found it 1 place, but I know it hasn't been touched (tried disabling it once as said and everything went slower indeed)
BTW, as said my Voltage at Idle was 0.91V as yours before enabling Uncore OC Mode
When it now says 0.2V, which Value in Ryzen Master should I look at? Tell me the name from the Screenshots of Ryzen Master

Differences:

You have faster RAM
Your MB is same brand but costs twice as much so it would be weird if you don't get better results than me with the same settings (you do, I must have Uncore OC Mode Enabled to get same Blender Benchmark results)
Newer kernel (my experience = better performance with every version for Ryzen Gen 3)
As said you use the CoreFreq CPU driver(?)

olejon on 22 Jun 2020

I suppose

You're using the CoreFreq CPU driver, no?

Yes all P-State tweaks done with _CoreFreq_. But I've not programmed a governor for AMD yet. Thus OSPM is under the control of the Linux ACPI module.

You have CPB 2 places in your UEFI? Maybe I do too but haven't looked, just found it 1 place, but I know it hasn't been touched (tried disabling it once as said and everything went slower indeed)

Yes, the Crosshair is also showing this CPB issue, 2 places for the same name. It's confusing.

BTW, as said my Voltage at Idle was 0.91V as yours before enabling Uncore OC Mode

When it now says 0.2V, _which Value in Ryzen Master should I look at?_ Tell me the name from the Screenshots of Ryzen Master

I believe it is called CPU Voltage under Voltage Control

Differences:

You have faster RAM

Your MB is same brand but costs twice as much so it would be weird if you don't get better results than me with the same settings (you do, I must have Uncore OC Mode Enabled to get same Blender Benchmark results)

Newer kernel (my experience = better performance with every version for Ryzen Gen 3)

As said you use the CoreFreq CPU driver(?)

Perhaps but the Processor P-States remain what they are, whatever the bells and whistles of the motherboard.
My goal here is to control the P-States: next steps will be to set some manual frequency ratios in BIOS then dump the registers changes within CoreFreq driver to understand how BIOS is _stabilizing_ frequencies...

cyring on 22 Jun 2020

You mean the one at bottom right? Saying V[0.20]. Is that bad or anything?
I'm not into Voltage tuning AT ALL.
I haven't set ANY integer or string manually, only Drop-Downs. Only the D.C.O.P. changes automatically what you already know:

RAM voltage to 1.35

RAM timings, for me 16-18-18-36

RAM frequency for me 3200 MHz

Yours show V[0.91]. What's the diff?

OK, so it's a _CoreFreq_ bug
Adding it to the todo list

Checking code, threshold is 0.15V and it applies only to the lowest voltage when reading falls down to zero.
Also threshold is not applied to current value (middle column)
So it might not a bug and the voltage has been read at 0.2 for some reasons.
https://github.com/cyring/CoreFreq/blob/ab750ba6df881c4a8e959797a23689800f54ce19/corefreq.h#L492

cyring on 22 Jun 2020

Well I'll check later what Ryzen Master says with current config totally idle
Then I'll Disable Uncore OC Mode and the AI Tweaker > PBO option and recheck RM and CoreFreq (surely goes back to 0.91V at idle)
Not like I need the extra speed in day to day use
BUT a good way to:
- Confirm that my Noctua fan holds its promise about best turbo/overclocking headroom, using only 1 of the fans
- Check that Linux beats Windows also when overclocked
- Fun to push HW to its limit when only danger is that it does a soft shut off (hasn't happened, won't push more as typical for Intel where you kind of have to since stock sucks)

olejon on 22 Jun 2020

As seen in Screenshot, CoreFreq is NOT reading the wrong Voltage. Ryzen Master also goes down to 0.2V in when Idle, although, since it's Windows, it goes much more up and down, but definitely in the 0.2V area many times during 1 minute
When reverted to Uncore OC Mode = Disabled and AI Tweaker > PBO = All Auto, CoreFreq again shows V[0.91]
BUT: Conic Compute > Hyperboloid of two sheets doesn't even manage to bring the CPU Temperature to 70°C (stays in the 60's)
Voltage goes to 1.07V during that stresstest
So why lose performance? Now it's just overkill cooling and CPU & PSU can handle the OC it seems
I have a Wattmeter I can plug in between wall outlet and PSU and see how much the PC in total uses during a Blender Benchmark. Would be interesting to see how close to the PSU's 500W it comes

olejon on 22 Jun 2020

The 0.20V screenshot was a "lucky one":

Voltage changes like this, in order of most frequent reading:
1 0.91V
2 0.96V
3 0.20V

So it's kind of like Ryzen Master, only that RM shows way more levels between 1-3 (and many more decimals, not that that matters).

I'm back to the OC mode. One can set a maximum temperature in AI Tweaker > Precision Boost Override too say avoid it never reaches 90°C, which I guess is when the Orange Color shows, but I have never reached above 89°C, as said, and that was just briefly before down again to 87-88, and only during Blender Benchmark fishy_cat.

olejon on 23 Jun 2020

I tend to check the ASUS website if there is a BIOS update. Saw there was one from June 17 published. Probably one for you too?

After installing it, Linux boots fine, but see these and journalctl shows the IRQ stuff in Red

juni 24 00:52:21 tux kernel: do_IRQ: 1.55 No irq handler for vector
juni 24 00:52:21 tux kernel:   #2
juni 24 00:52:21 tux kernel: do_IRQ: 2.55 No irq handler for vector
juni 24 00:52:21 tux kernel:   #3
juni 24 00:52:21 tux kernel: do_IRQ: 3.55 No irq handler for vector
juni 24 00:52:21 tux kernel:   #4
juni 24 00:52:21 tux kernel: do_IRQ: 4.55 No irq handler for vector
juni 24 00:52:21 tux kernel:   #5
juni 24 00:52:21 tux kernel: do_IRQ: 5.55 No irq handler for vector
juni 24 00:52:21 tux kernel:   #6
juni 24 00:52:21 tux kernel: do_IRQ: 6.55 No irq handler for vector
juni 24 00:52:21 tux kernel:   #7
juni 24 00:52:21 tux kernel: do_IRQ: 7.55 No irq handler for vector
juni 24 00:52:21 tux kernel:   #8
juni 24 00:52:21 tux kernel: do_IRQ: 8.55 No irq handler for vector
juni 24 00:52:21 tux kernel:   #9
juni 24 00:52:21 tux kernel: do_IRQ: 9.55 No irq handler for vector
juni 24 00:52:21 tux kernel:  #10
juni 24 00:52:21 tux kernel: do_IRQ: 10.55 No irq handler for vector

...

juni 24 00:52:21 tux kernel: ACPI BIOS Error (bug): Failure creating named object [\_SB.PCI0.GPP8._DSM], AE_ALREADY_EXISTS (20190703/dswload2-327)
juni 24 00:52:21 tux kernel: ACPI Error: AE_ALREADY_EXISTS, During name lookup/catalog (20190703/psobject-221)

Any ideas?
The BIOS update seems to be AMD only stuff. You know what this is?
After install it said it was updating the "Led firmware" and nothing more before booting
Changelog:

PRIME X570-PRO BIOS 2203
Update AMD AM4 AGESA PI 1.0.0.1

olejon on 23 Jun 2020

I tend to check the ASUS website if there is a BIOS update. Saw there was one from June 17 published. Probably one for you too?

Thanks for the tips

After installing it, Linux boots fine, but see these and journalctl shows the IRQ stuff in Red

juni 24 00:52:21 tux kernel: do_IRQ: 1.55 No irq handler for vector
juni 24 00:52:21 tux kernel:   #2
juni 24 00:52:21 tux kernel: do_IRQ: 2.55 No irq handler for vector
juni 24 00:52:21 tux kernel:   #3
juni 24 00:52:21 tux kernel: do_IRQ: 3.55 No irq handler for vector
juni 24 00:52:21 tux kernel:   #4
juni 24 00:52:21 tux kernel: do_IRQ: 4.55 No irq handler for vector
juni 24 00:52:21 tux kernel:   #5
juni 24 00:52:21 tux kernel: do_IRQ: 5.55 No irq handler for vector
juni 24 00:52:21 tux kernel:   #6
juni 24 00:52:21 tux kernel: do_IRQ: 6.55 No irq handler for vector
juni 24 00:52:21 tux kernel:   #7
juni 24 00:52:21 tux kernel: do_IRQ: 7.55 No irq handler for vector
juni 24 00:52:21 tux kernel:   #8
juni 24 00:52:21 tux kernel: do_IRQ: 8.55 No irq handler for vector
juni 24 00:52:21 tux kernel:   #9
juni 24 00:52:21 tux kernel: do_IRQ: 9.55 No irq handler for vector
juni 24 00:52:21 tux kernel:  #10
juni 24 00:52:21 tux kernel: do_IRQ: 10.55 No irq handler for vector

...

juni 24 00:52:21 tux kernel: ACPI BIOS Error (bug): Failure creating named object [\_SB.PCI0.GPP8._DSM], AE_ALREADY_EXISTS (20190703/dswload2-327)
juni 24 00:52:21 tux kernel: ACPI Error: AE_ALREADY_EXISTS, During name lookup/catalog (20190703/psobject-221)

Any ideas?
The BIOS update seems to be AMD only stuff. _You know what this is?_
After install it said it was updating the "Led firmware" and nothing more before booting
Changelog:

PRIME X570-PRO BIOS 2203
Update AMD AM4 AGESA PI 1.0.0.1

Have you clear the CMOS right after the update ?
If not, you have garbage values in the _nvram_

cyring on 24 Jun 2020

The BIOS always resets to default values after an update and I go through all the settings again and set them as before (and check if new ones), instead of loading a saved profile.

Hm. Maybe a total power off then (PSU too). If not, does a hard CMOS reset ("short circuit" pins) do more than resetting the BIOS in UEFI, really? Fixing NVRAM shit is typically a Mac thing. Think many knows Option + Command + P + R lol (and the SMC reset one).

Had something similar, don't remember when, if was switching to Ryzen or the AMD GPU, PCIe3 to 4 NVMes... That was fixed quickly by an Ubuntu update.

olejon on 24 Jun 2020

The BIOS always resets to default values after an update and I go through all the settings again and set them as before (and check if new ones), _instead of loading a saved profile_.

Hm. Maybe a total power off then (PSU too). If not, does a hard CMOS reset ("short circuit" pins) do more than resetting the BIOS in UEFI, really? Fixing NVRAM shit is typically a Mac thing. Think many knows Option + Command + P + R lol (and the SMC reset one).

Had something similar, don't remember when, if was switching to Ryzen or the AMD GPU, PCIe3 to 4 NVMes... That was fixed quickly by an Ubuntu update.

It still written today in the ROG manual to clear the BIOS _RAM_ after an update. To my knowledge, nothing is automated. A jumper should be on board ?

cyring on 24 Jun 2020

Well tried that... The jumper actually resets less than the BIOS update... Any BIOS update does the same as Load optimized defaults (F5)
With the jumper it's immediately clear when I see my favorites in UEFI they're still there. After F5 + Save, I see that short circuiting the jumper actually keeps some AMD RAM stuff (they're not cleared by jumper just UEFI (F5)
I always update the BIOS over the Internet in UEFI and it always says Update successful. System will Reset! (and must set all settings again, or load previous profile but I never use older profiles on new BIOS)
I took out the RAM chips for a while, for another reason
Still bug. It's clearly a BIOS bug, kernel says so... (Windows maybe too if digging in logs)

juni 24 04:02:08 tux kernel: ACPI BIOS Error (bug): Failure creating named object [\_SB.PCI0.GPP8._DSM], AE_ALREADY_EXISTS (20190703/dswload2-327)
juni 24 04:02:08 tux kernel: ACPI: [Firmware Bug]: BIOS _OSI(Linux) query ignored

Meh, unless I notice any regression. Assume this will be fixed in the not too distant future

olejon on 24 Jun 2020

Some ACPI Device Specific Method _DSM linked to GPP8 in System Bus _SB

Search for GPP8 in whole /sys/devices/LNXSYSTM
Mine is located at :

cat /sys/devices/LNXSYSTM:00/LNXSYBUS:00/PNP0A08:00/device:0f/path
\_SB_.PCI0.GPP8

cyring on 24 Jun 2020

Nothing...

/sys/devices/LNXSYSTM:00# find ./ -name "*GPP8*"

Suppose there's no point in reinstalling the linux-firmware amd-microcode packages. Will check with Ubuntu 20.04 Live USB if also there. Just in case.

Think my MB has BIOS Flashback to revert, but I've never used such a feature. Next time I'll google properly before updating. The April BIOS was good.

Did you understand the changelog? Have you installed it successfully?

olejon on 24 Jun 2020

Nothing...

/sys/devices/LNXSYSTM:00# find ./ -name "*GPP8*"

Suppose there's no point in reinstalling the linux-firmware amd-microcode packages. Will check with Ubuntu 20.04 Live USB if also there. Just in case.

Think my MB has BIOS Flashback to revert, but I've never used such a feature. Next time I'll google properly before updating. The April BIOS was good.

Did you understand the changelog? Have you installed it successfully?

I won't install any BIOS. All features are working fine

Since Zen has been introduced, ACPI tables have been an issue. See the kernel history....

ROG CROSSHAIR VIII HERO had been released with:

BIOS 1201 11/18/2019

but it has however a remaining log trace ...

ACPI: [Firmware Bug]: BIOS _OSI(Linux) query ignored
ACPI BIOS Error (bug): AE_AML_PACKAGE_LIMIT, Index (0x000000007) is beyond end of object (length 0x6) (20200326/exoparg2-393)
ccp 0000:0c:00.1: ccp: unable to access the device: you might be running a broken BIOS.

... which is fine for _CoreFreq_ because I'm working on my own [P][C]-States sub-drivers for Zen

My EFI entry tries to blacklist any of them:

title   Arch Linux
linux   /EFI/Linux/vmlinuz-linux
initrd  /EFI/Linux/amd-ucode.img
initrd  /EFI/Linux/initramfs-linux.img
options root=/dev/disk/by-label/root rw quiet break=n add_efi_memmap nmi_watchdog=0 selinux=0 loglevel=3 rd.systemd.show_status=auto rd.udev.log-priority=3 consoleblank=0 vt.color=0x03 modprobe.blacklist=pcspkr,nouveau,k10temp,acpi_cpufreq idle=halt cpu0_hotplug audit=0

which results in _CoreFreq_:

Linux:                                                                          
|- Release                                                       [5.7.4-arch1-1]
|- Version                      [#1 SMP PREEMPT Thu, 18 Jun 2020 16:01:07 +0000]
|- Machine                                                              [x86_64]
Memory:                                                                         
|- Total RAM                                                         32856504 KB
|- Shared RAM                                                          117756 KB
|- Free RAM                                                          30792460 KB
|- Buffer RAM                                                           79832 KB
|- Total High                                                               0 KB
|- Free High                                                                0 KB
CPU-Freq driver                                               [         Missing]
Governor                                                      [         Missing]
CPU-Idle driver                                               [         Missing]

Missing being the sign that nothing is in control, letting me choose Target P-States
For example, decrease from default 35 to 22 but put back CPU #12 to 35
2020-06-24-103525_644x1012_scrot

But target ratios are still _moving_ on their own, resulting in
CPU #12 = 36 and its SMT CPU #28 jumping between 43, 36
other CPUs are also not fully in control.

I believe its part of the collaborative stuff between Processor and some software, feeding back to it the requested Target. So far studying those entry functions: acpi_cpufreq_fast_switch , acpi_idle_enter and cppc_cpufreq_set_target

cyring on 24 Jun 2020

I'm noticing this strange behavior

with acpi_cpufreq kmod blacklist and thus no cppc enabled (to my understanding)
_CoreFreq_ started with AutoClock=0 argument to read a BCLK of 100MHz
based on the latest develop of _CoreFreq_ which has been changed to display 3 items in the Target selector:
current performance ratio
current performance frequency
programmed target ratio

The frequencies plan as follow:

|- Base Clock                                                          [100.000]
|- Frequency            (MHz)                      Ratio                        
                 Min    400.00                    <   4 >                       
                 Max   3500.00                    <  35 >                       
|- Factory                                                             [100.000]
                       3500                       [  35 ]                       
|- Performance                                                                  
   |- OSPM                                                                      
                 TGT   3500.00                    <  35 >                       
|- Turbo Boost                                                         [ UNLOCK]
                 XFR   4700.00                    [  47 ]                       
                 CPB   4600.00                    [  46 ]                       
                  1C   2800.00                    <  28 >                       
                  2C   2200.00                    <  22 >

>> meaning 3 enabled Ratio P-States: P0 = 35 , P1 = 28 and P2 = 22

Now stressing one Core, at least one physical, preferably the _best_ one which is CPU#0
set all Target to lowest ratio P-state, value P2 = 22
set its SMT Core, CPU#16 , to ratio P1 = 28
>> the obtained frequency is 2.8GHz
but now set CPU#16 to P0 = 35

>> it triggers Turbo Boost !
It's a strange behavior which mimics the Turbo conditions found in the Intel Nehalem architecture

cyring on 24 Jun 2020

Luckily it was easy to revert back to BIOS 1407 (April). It did enhance performance somewhat (not reverting but upgrading to that one from the previous).

Load optimized defaults (F5) in UEFI and save
Boot Windows (could've used Linux)
Format USB drive to FAT32
Download that BIOS file and put it on the USB
Opening the EZ Flash tool in UEFI
It automatically switched to the disk with the .CAP file
Flash it, and it says Update successful, System will Reset!
Back to normal
Manually go through UEFI settings again (right now all stock)

Your BIOS version is the one AMD specifically asked people to install, fixing the "ABBA bug", after installing 1005 first, to fix PBO really not doing anything on many Ryzen 3rd Gen CPUs
Indeed the manual says that one should not update the BIOS if the system is stable, ALL manuals for any brand/arch say this of course
But my journey has shown improvements with every update, like benchmark performance and PCIe4 improvements, maybe also allowing modules like k10temp reading the temperatures, but that I don't know as that update came more or less at the same time as a BIOS update
Also systemd-fstrim working finally, after not working the first half year or so (and can't be trimmed on Windows as no NTFS)
IIRC both AMD and Corsair employees recommended to install all the 2019 BIOS updates, at the time I looked. Corsair to get the best from their new PCIe4 NVMes. Then using "Corsair Toolbox" on Windows to update the firmware. Only available on Windows, but didn't cause any harm to my NVMes which only has Linux on them
There are 2 more 2019 updates than yours, 1404 & 1405

Changelogs for those:

1404 (2019):

Update AM4 combo PI 1.0.0.4 patch B
Support Ryzen™ 2000-series APU
You will not be able to downgrade your BIOS after updating to this BIOS version

1405 (2019):

Improve system performance

1407 (April 2020 one I run now):

Improve system stability
Add BCLK Frequency and SB Clock Spread Spectrum items into BIOS Advanced mode.

I remember now what caused some kernel errors before (without problems). Was The radeon driver (IDK after system or BIOS update), but as said that was fixed within days by Ubuntu kernel and/or grahpic stack devs
One thing I noticed with this June BIOS 2203 update was Ryzen Master actually showing all cores sleeping when system is idle, so some CPU stuff in it
Why the June BIOS has such a drastic jump in BIOS version IDK, up to 2203. Only thing is Update AMD AM4 AGESA PI. Staying on 1407 until I see some more interesting changelog. Not installing it if it says one cant downgrade after updating. Also waiting at least 1 month and google first (Linux + 3950X)
Regarding you seeing the CPU "denying" to adhere, I'm pretty sure I read many places after launch + YouTube videos that the Ryzen 3rd Gen CPUs have a lot of protection built-in. If they live in the BIOS, Motherboard or the CPU itself I don't know
Basically have to use the UEFI tweaker, like setting Core Ratios for all to say 4200 MHz when turbo (ratio 42), but there's little headroom there without very good cooling (but safety should just turn off the system if too much OC, my experience for the 3600X), as it should, people should NOT have to fiddle like one basically must on Intel i9 CPUs to achieve the specs

olejon on 25 Jun 2020

Glad you successfully fallback to previous firmware and thank you for this detailed procedure.
As I told you I'm not in a hurry to upgrade BIOS. There are still so many _bits_ I have to discover and to program into _CoreFreq_

Using the latest version, are you also noticing than for a standard target performance ratio of 35, processors answers with a value of 36 ?

To monitor this, press key ! to toggle to absolute frequency

Are you also observing that 36 is the _magic_ value which triggers Turbo Boost ?

cyring on 25 Jun 2020

And guess what? JUST about 24 hours later came a full kernel stack update + updates to some core libraries like libc not shown here, which may have a fix for the bug :man_shrugging:
Although say Ubuntu 20.04 has a newer kernel all bug fixes and performance updates are backported. And about 3 months after 20.04.1 or 20.04.2 has been released, 18.04 will get the 20.04 HWE stack, kernel and graphics, running the same
But won't update to the June BIOS again to check heh
Package names don't reflect kernel version, it's 5.3.18, but these came:

linux-image-5.3.0-61-generic linux-modules-extra-5.3.0-61-generic linux-headers-5.3.0-61-generic linux-headers-5.3.0-61 linux-modules-5.3.0-61-generic

Can NOT test with a Live USB even with kernel hotswap as a reboot is required to see the bug-entries
But yeah, after downgrading to 1407, they're gone now. BTW only thing showing in plain text after that update was also Updating Led Firmware, but that might have been really a downgrade from the June one
As to your question, nope, I get the exact same values as the freq BASH alias I have on all my devices, which runs this command:

watch -n 2 cat /sys/devices/system/cpu/cpu*/cpufreq/scaling_cur_freq

Now I can't say this isn't because of me running a newer BIOS than you or not, as per changelogs having performance and stability improvements you don't have, as I've not used the ! toggle since fall 2019 and DON'T remember how it was back then
Screenshot:

olejon on 25 Jun 2020

Here already running one of the latest kernel version 5.7.5 but still having ACPI issue.
My BIOS might be not ACPI tables compliant or the Kernel still don't know how to process some of them.
Your P-State ratio is showing 22 because, I believe, your processor is under the control of acpi_cpufreq.
In my use case, I have blacklist any kernel modules, to study P-States in a deterministic way

Without CPUFreq drivers (confirmed by Missing) I can program things these ways:

You program ratio to 22 then you stress CPU and its frequency ratio is capped to22 (as expected)
The same happens, if you program the P-State to 28
But as soon as you program 35, idle shows 36 and full load goes up to 47

To summarize: I don't get the 47 ratio boost with P-States 22 and 28 but only with the P0 P-State of factory value 35

cyring on 25 Jun 2020

This is what we get from the Rocket SSD
ROCKET NVMe 4 0 WITH HEATSINK

cyring on 26 Jun 2020

CrystalDiskMark is OK, sometimes, but can be wildly inaccurate too
Manufacturers use ATTO Disk Benchmark when setting their specs

From Tom's Hardware:

ATTO is a simple and free application that SSD vendors commonly use to assign sequential performance specifications to their products. It also gives us insight into how the device handles different file sizes

Straight from Corsair:

ATTO Disk Benchmark has long been an industry staple for measuring drive performance

It's just a form to get it from the official website (would not trust all the freeware sites) where you can use false name and whatever test email, and newsletter is OFF by default and I've never gotten any email afterwards
My download link below should work. Just installs 3 files in same folder of your liking. Start Menu icon can be deleted of course, no desktop icon is added and uninstall is as easy as any, and zero bloatware...
Link you get is generic and should work: https://www2.atto.com/asset/1:atto-disk-benchmark
Made an NTFS partition at the end of my /home NVMe. Don't remember if it's the CPU bound or Chipset bound, but they get the same speed in Linux so...
I just use these options to best match the Linux Benchmark, but you can try the default full run. For me it is unnecessary, I see these results after reaching the 128 KB test

Windows - Corsair MP600 - Basically the first proper PCIe4

olejon on 26 Jun 2020

I fear that WSL won't offer more than the TSC
WiP #191

cyring on 26 Jun 2020

Oh lol, I was about to try CoreFreq on the Windows Linux subsystem one day, but assumed it wouldn't work, but worth a shot since it's easy enough to try. That if I could actually install the build-essential metapackage. Otherwise I would assume not. Also noticed screen only works as root/sudo. Don't like to run e.g. apt (or any package manager) to install/update many packages without screen or similar, on any system.

Kind of odd Microsoft has added that option. First they kind of had their subsystem where you could run BASH directly from the Start Menu (through search only I think, was a hidden link) or Shift + Right Click in a folder in Windows Explorer. But they removed that without saying really. BASH just exited.

But was no good anyways when actually worked. Lacked so much. Ubuntu from the Microsoft Store runs the latest LTS, or you can choose a compatible LTS. Then it works, and has more or less what macOS has by default. Easier to install more, though. Not Homebrew and Apple making that harder or screwing up installed packages for every version.

It's weird to have cmd.exe which is horrible but can do much more than one might think. Many driver packages, some software and enterprise AD use it for setup still... Has tons of AD/networking commands but very undocumented. Unless you've taken a MS certification course or something, it's hard to work with.

Powershell I played with a little. Long story short, even our Powershell expert couldn't make it do what a BASH script I wrote in 30 seconds did (read webcam barcode value and open browser to X URL with Y parameters).

So now Window's got 3! They gotta keep cmd for drivers and certain software, but hide it and just use BASH... IMHO.

Pretty nicely integrated with your / Linux folder right there in Windows Explorer and everything. Little like Chromebooks but they also run GUI software just fine like GIMP etc.

Like... Have you tried just to get the MD5 value of a file on Windows without a 3rd party program? Found some official Microsoft package to add an insanely long command to cmd to do checksums but turned out to be deprecated, without saying any option... Now that's easy. And much more basics that a freakin' OS should be able to. Lots of Windows packages does "Verifying download/install" but IDK what mechanism. Apple has md5 (not called md5sum) and shasum like Linux, but doesn't use BASH by default anymore.

olejon on 26 Jun 2020

Is Hyper-V the Hypervisor of WSL-2 ?
Booting the _CoreFreq_ ISO into a Hyper-V VM behaves the same as within WSL.
Thus is WSL a particular VM of Hyper-V, but without showing up into the VM manager.

Many MSR can be queried. Although they return zero, they don't crash. They are thus not pathrough, but trapped.

But APERF/MPERF counters, which are similar to MSR, crash. I feel they are trapped by Hyper-V, but something is missing.
Probably, some settings are missing in my WSL configuration, such as Xen, KVM and VBOX let you define rules when a MSR is called.

Why Hyper-V makes possible a virtualized TSC, and not APERF /MPERF ?
Especially when: virtualizing the TSC, the hardest part, has been achieved.

cyring on 26 Jun 2020

Well one search and...

Does WSL 2 use Hyper-V? Will it be available on Windows 10 Home?
WSL 2 will be available on all SKUs where WSL is currently available, including Windows 10 Home. The newest version of WSL uses Hyper-V architecture to enable its virtualization. This architecture will be available in the 'Virtual Machine Platform' optional component. This optional component will be available on all SKUs. You can expect to see more details about this experience soon as we get closer to the WSL 2 release.

FAQ: https://docs.microsoft.com/en-us/windows/wsl/wsl2-faq

olejon on 26 Jun 2020

Definitively it's the same Hypervisor: _CoreFreq_ is running in an ArchLinux VM and we're reading the same Vendor ID
CoreFreq_HYPER-V

cyring on 26 Jun 2020

It's been so hot in Europe these last days, I tuned my CPUs to cool down: No OC, especially PBO

cyring on 27 Jun 2020

Introduction

Yeah, also very hot In Norway, for weeks :sun_with_face:

I wanted to get to the bottom of which UEFI Settings give better performance
For this I used Windows (Idle) + Blender Benchmark classroom + Ryzen Master to monitor
I compare 2 findings with Linux (Idle) using Blender Benchmark classroom + CoreFreq to monitor
For this I did reset all the UEFI Settings to default, then set:
- Logically D.C.O.P = Enabled to actually use my RAM chips' speed
- Logically CSM = Enabled to be able to boot Windows, since legacy boot

NOTE: I divide this into 2 Categories: AMD OC found under AMD Overclocking and Manual OC found under AI Tweaker, or for you Extreme Tweaker. From what I see they have the same Settings. One (fancy manual) setting was added in a later BIOS version than yours. I've never touched it.

Windows Screenshots. Every screenshot has bullet points above = Which UEFI Settings were changed
Linux Comparison Screenshots
UEFI Screenshots of the Settings
CONCLUSION: Temperatures, Watts and Other Observations

Windows - Run 1

Nothing changed

Windows - Run 2

In Ryzen Master: OC Mode set to Auto Overclocking and Reboot
This changes AMD OC only
Change approx. from 3850 to 3920 MHz
Speed gain approx. 70 MHz
RESULT: 1 Second better than Run 1

Windows - Run 3

Also RAM Gear Down Mode = Disabled
This changes AMD OC only
Default is Enabled
This setting can be seen in Ryzen Master as GearDownEn = Enabled/Disabled
RESULT: 3 Seconds better than Run 1, meaning this Simple Setting actually helps performance as said in videos, at no cost(?)

Windows - Run 4

Also Power Down Enable = Enabled
This changes AMD OC only
Default is Auto, may be affected by Gear Down Mode = Disabled, I did not check, waste of time
This setting I can not find in Ryzen Master
RESULT: No effect. Useless!

Windows - Run 5

This is "Manual OC"
Reverted the useless back to Power Down Enable = Auto
In AI Tweaker, set Precision Boost Override = Enabled
Let the 3 others stay at Auto
RESULT: 14 Seconds better than Run 1

Linux Clone of same Settings as in Run 3

RESULT: 24 Seconds better than comparison

Linux Clone of same Settings as in Run 5

RESULT: 22 Seconds better than comparison

UEFI `AMD OC` Built-In PBO Settings

NOTE: Both these set by Ryzen Master
First Screenshot sets PBO to a hypothetical 100 MHz
Second Screenshot simply changes Overclock = Enabled while the rest stays at Auto
RESULT: Peak MHz about 70 MHz higher on Windows

UEFI `AMD OC` Built-In RAM Settings

NOTE: Gear Down Mode = Disabled gave better Blender Benchmark result

NOTE: Power Down Enable = Enabled default is Auto, as said may be affected by Gear Down Mode

UEFI `Manual OC` Settings

Have only tried this one of all the stuff in AI Tweaker, and don't plan to mess with others
As said in another post, it is NOT necessary to enable Uncore OC Mode (my bad) which is under AMD OC Settings
Only have to set Precision Boost Override = Enabled and leave rest at Auto
No reason to set 100 MHz here too
Not only does it increase Peak MHz by approx. 200 MHz on both Windows and Linux
It also seems to make the Peak MHz more stable, in both Ryzen Master and CoreFreq
By more stable I mean when you look at the MHz of all the Threads they fluctuate less
Help text is pretty clear. Allows the "CPU to run (beyond its X defined limits) to the limits of the board", sounds OK for the Motherboard... Seems scary to you?
CPUs have their own protections, and several, no?
I think my Skylake simply shut off when too much OC. That's why I had this sensor software to reboot it if >= 96°C in case, so server services would continue to work
My previous 3600X simply rebooted to POST and just pressed F1 to revert a test OC I did I more or less knew wouldn't work since it had the stock cooler. I may be wrong and it is the other way around. Hasn't happened with this one so IDK. Per now this Not Enabled before I read more about it, FYI

CONCLUSION

The following observed when running Blender Benchmark classroom
When Ryzen Master's OC Mode = Default then PPT = 95-97% of 142 Watts
When Ryzen Master's OC Mode = Auto Overclocking then PPT = 100% of 142 Watts, showing Orange Color, but since 197 Watts+ is apparently no problem, it's just an indicator for that PPT
When run in Linux, CoreFreq shows a pretty darn stable 142 Watts too
Guess it is a "cap", meaning if PPT, TDC or EDC reach 100%, the CPU does not give more performance
Temperature on Windows then = 65-67°C, but NOT valid, fans barely spin up at that with my current UEFI fans config
Temperature on Linux then = 64-66°C, but NOT valid, fans barely spin up at that with my current UEFI fans config
POINT IS: No significant difference seen in Temperature between Ryzen Master OC Mode Default and Auto Overclocking
Setting Gear Down Mode = Disabled did not affect Temperature either. This was as said with OC Mode = Auto Overclocking so PPT = 100% of 142 Watts, still the better result
Remember for the following, this is what I saw during the Blender Benchmark, but all in all I think I have seen Windows give a higher Peak MHz, without "winning" because of it
The effect of OC Mode = Auto Overclocking is as said a Peak MHz of approx. 70 MHz higher, but it is higher on Windows actually
On Windows OC Mode = Auto Overclocking causes 3850 MHz to 3920 MHz (more or less)
On Windows OC Mode = Auto Overclocking causes 3835 MHz to 3905 MHz (more or less)
In this mode NO stresstest gets it up to 70°C even with 30°C+ ambient, generally it Idles at 3-5°C above ambient so all good there
Forgot to pay attention to Voltage, but Conic Compute > Hyperboloid of two sheets = 1.09 V

Now we set Precision Boost Override = Enabled under Manual OC
Still OC Mode = Auto Overclocking and Gear Down Mode = Disabled for best results
PPT changes from 142 Watts to 395 Watts
TDC changes from 95 A to 255 A, but it never reaches that of course, always stays at Green Color
EDC changes from 140 A to 255 A, but it never reaches that of course, always stays at Green Color
During run, Ryzen Master shows PPT = 50% of 395 Watts = 197.5 Watts, and totally stable at that
CoreFreq shows a bit more fluctuation, jumps between 193-203 Watts, but not constantly, so an average approx. Windows
I can only guess there is some "cap" around 203 W on Linux with this particular OC, while Windows doesn't want to go over 197.5 W, it doesn't go below either, while Linux does (some balancing?)
Temperature on Windows then = 85-86°C
Temperature on Linux then = 84-85°C
BTW: Temperature is not the cap. I've seen 88°C in scene fishy_cat in this mode on Windows, MHz and everything totally stable, but Watts is lower with that scene
Forgot to pay attention to Voltage, but Conic Compute > Hyperboloid of two sheets = 1.28 V

DESPITE THIS: No worries when the Winner of all CPU Benchmarks run to date here is Linux. And that from a year ago, when I bought the 3600X
Numbers are interesting but never tell the full story between 2 very different Operating Systems. Despite stable Peak MHz indicate Windows should win, it never does...

EDIT: Would be cool if CoreFreq could read PPT, TDC and EDC MAX values as Ryzen Master shows them. Well at least PPT since CoreFreq shows Watts. Possible?
I don't expect any long answer to this. Just a test I wanted to do and why not share it when we have the same CPU now :-)

olejon on 27 Jun 2020

Awesome.

PBO is indeed helping a lot in performance score. But did you mention which policy plan is in used: Windows default or AMD policy ? The last is keeping the frequency target to a high value, at the expense of temperature; whereas Windows Balanced mode is keeping frequency among the lowest P-State ratios.

Would be cool if CoreFreq could read PPT, TDC and EDC MAX values as Ryzen Master shows them. Well at least PPT since CoreFreq shows Watts. Possible?

One software on GitHub is getting those values by reversed engineering.
I've not try it yet.
And specs just give the basic information. Not enough to program a safe algorithm.

cyring on 27 Jun 2020

Impossible to replicate Run 3 on Windows

After having reset all UEFI Settings I found Linux performed the same as Run 3 with only D.C.O.P.
So I had to check Windows, and it did a little worse, so seemed like Windows wanted the Run 3 Settings
The Ryzen Master stress test for OC Mode: Default reached 97% of PPT (142 W) as before

But first the Voltages:

Windows:
- Running classroom: 1.14 - 1.16 V
- Idle = 0.32 V (up some times, but for the vast majority of time)
Linux:
- Running classroom: 1.06 - 1.09 V
- Idle: 0.20 - 0.91 - 0.96 V (vast majority of time at 0.91 V, but all cores show they've been at 0.20 V too, and occasionally goes 0.96 V)

Frequencies after having run 1 minute:

Windows: 3825 - 3850 MHz
Linux: Approx. 15-20 MHz lower, never hitting 3850 MHz again after the very beginning

ACTUAL RESULTS:

So I used Ryzen Master again to set OC Mode: Auto Overclocking and Reboot
The Ryzen Master stress test after Reboot then showed 100% of PPT (142 W) as expected
Reboot and let Idle
Despite this, time went up with a couple of seconds
So I rebooted again and set Gear Down Mode = Disabled too
Despite this, time went up with a couple of seconds
I think first Ryzen Master showed some 95% of PPT = 135 Watts
But this went down to 93% of PPT = 132 Watts
Results for classroom went like this: 4m23s, 4m26s, 4m29s, 4m30s, and then I facepalmed and gave up
Booted Linux and was afraid it would perform worse there too, for some reason...
Linux, as before, ALWAYS uses about 96% of PPT, a little above 136 Watts
BUT nope, Linux did just as always
Result for classrom: 4m plus/minus 1 second, every time
And this result it gets as said in the beginning with only D.C.O.P.
The other UEFI settings I change do not have anything to do with OC

POSSIBLE REASONS:

It depends on Windows' mood or time of day? :man_shrugging:
The Chassis HDD Led does NOT blink for NVMes, and I see some blinking during run on Windows, as it is on a dedicated SATA SSD, so maybe Blender Benchmark depends on I/O speed after loading the scene?
You have both OS' on the same NVMe so you can test that
BUT: That still does not explain why Linux beats Windows by a lot in Geekbench

TO YOUR QUESTION:

I use the Balanced Power Plan as it's the only one really made for Windows 10. Changing to another removes a Performance Slider from the new Settings App
It acts just like ondemand on Linux when I see the frequencies using HWMonitor (same as ! in CoreFreq)
The Windows Built-In Power Plans can be reset to default
3rd Party Power Plans can not, you must reinstall the whole package that installs them...
If you open the ADVANCED DETAILS of a Windows Power Plan you see all the settings, including after how many minutes the monitor(s) should turn off and all that
And if you see the Windows Built-In Performance plan compared to the Ryzen Performance plan they are pretty identical
The Ryzen Performance does nothing more than make the CPU run at Base Clock MHz all the time
This can be achieved in Linux using sysfs too... Setting governor to performance or a Specific Frequency for all 32 in sysfs (think this was needed, changing governor alone didn't work, it still boosted)
I made a BASH script to toggle it when I had the 3600X, it does NOT improve benchmarks
At best, IF the system is Idle, you get a tiny little advantage when opening something that is not in RAM... Except NVMes...
I have tried the Performance Slider in the new Settings App, without seeing any difference, IIRC it's just 3 steps, basically a replacement for the old Power Saver, Balanced and Performance. Those old ones found in the "old" Control Panel are just there for legacy reasons it seems, considering the Slider disappears if not using Balanced as said, at least it was like that, and I'm 100% sure it disappears if you choose a 3rd Party Plan
BUT remember, all those Windows Vista-Era plans have the same settings, if you look into them... So you can set the Balanced plan to the same settings as seen in Ryzen Performance, and IDK what the Slider is for then, maybe it disappears
And basically just 2 settings have to do with the CPU, the one called something like "Active/Passive Cooling" and the one where you set Minimum/Maximum % of operation, which by default is 5% in all plans except the Performance ones, and what is 5% exactly? They only switch Minimum to 100%, Maximum is already 100% of course. That's is, IIRC

olejon on 27 Jun 2020

It depends on Windows' mood or time of day?

There are so many Windows service processes of any kind running that it's hard to measure and reproduce benchmark.
Even RM has a 100 to 300 MHz CPU overhead for its own monitoring.
Either Windows is stripped of any tasks to make it bare metal, or Linux Kernel is built with most drivers and security removed, such as page randomization. Then you will set processor registers to your will and barely get a reproductible and _understandable_ score.

In BIOS I have enable the Data-Fabric C-States but with no specs how to implement them in _CoreFreq_, I can't tell the benefits.

cyring on 27 Jun 2020

My Windows 10 is as Clean as it gets. All permanent or scheduled unnecessary services Disabled.

After boot:

Run Task Manager until The System Idle "process" is 99% constantly and no I/O (you never see 100%, that "process" has been there since Windows 95 or something)
So wait about 2-3 minutes
Open Ryzen Master and see that indeed 31 Cores are "Sleeping" all the time, sometimes 1 more wakes up for 1-2 seconds but it's always like that
I have of course run Blender Benchmark without Ryzen Master open and of course EXACT same results, the tiny 150-200 MHz it uses makes no difference for a heavy tool like Blender Benchmark
When I runch Geekbench I NEVER have anything open other than cmd.exe

I did try the Ryzen Performance Power Plan, and observed Ryzen Master for any changes:

No difference in Voltage
No difference in Watts
No difference in Amount of Cores "Sleeping", still 31 Cores sleeping for the most part
The "Peak Speed", although hard to observe, was maybe 40 MHz above, like instead of being mostly around 160 MHz it was more around 200 MHz, but I didn't bother sitting there staring for minutes
Remember just moving the mouse a little wakes up several Cores, so that Power Plan causes less CPU use than moving the mouse, if any at all
CONFIRMED: The only place the Ryzen Performance Power Plan is different than the standard Windows Balanced is that it sets "Minimum CPU % = 100", which does nothing as said above
The Ryzen Balanced Power Plan is even stupider. Its only difference is that it sets "Minimum CPU % = 99". Those using that are just stupid. No difference whatsoever in the values above

BTW: Not sure if you're ever going to find that the Cores actually Sleep. I've been watching and never seen a Core below 130 MHz - once it's below 130 MHz it says "Sleeping". So maybe "Sleeping" is just indicating very low MHz...

olejon on 27 Jun 2020

BTW: Not sure if you're ever going to find that the Cores actually Sleep. I've been watching and _never seen a Core below 130 MHz_ - once it's below 130 MHz it says "Sleeping". So maybe "Sleeping" is just indicating very low MHz...

But AMD says Processor has C-States. Those are mentioned in ASUS BIOS.
The Software Manual specifies the X86 MONITOR/MWAIT instructions to enter idle states.
This has to exist but we don't have so far the counters addresses to monitor them.

I'm reading you that RM may implement a CPU load minimum threshold to say: "it's sleeping"
But why such dummy algorithm ? RM is an AMD software: authors have full access to the NDA specifications to make an accurate software !

_CoreFreq_ version 1.79 for Hyper-V & WSL is released
You can download the ISO image to try under your Hyper-V manager
Or, just pull the last _CoreFreq_ source code, following WSL procedure to build with kernel headers at #191

cyring on 27 Jun 2020

The GPU vertical mounting _trend_

We it see more and more.
My case allows it.
I started googling for extender reviews and issues.
Lost in the comments jungle, I stopped at AMD

cyring on 27 Jun 2020

I'm reading you that RM may implement a CPU load minimum threshold to say: "it's sleeping"
But why such dummy algorithm?

It was just observation from RM. CoreFreq shows a MHz as low as 2 MHz here so not too worried about power consumption. Even if they have C-States does it mean totally "Sleeping", meaning 100% idle? Well your MONITOR/MWAIT may suggest that.

BTW: HWMonitor suggests no such idling. They stay at the approx. 2100 MHz last I checked, like ! in CoreFreq or reading from sysfs. Weird since that tool shows so much (from RAM temperatures to all fan speeds).

Question is if this happens on Linux anyways? Still waiting for official kernel cpufreq, cpuidle drivers. I seem to recall by searching or read it on Phoronix, that the kernel developers were at least working on a cpuidle driver (probably with AMD contributing).

Maybe even some Git version of the kernel has it already but must be enabled manually as it's still experimental?

Considering AMD's contributions to the Linux kernel I kind of expect those to officially come. Mind you the radeon driver on amd.com. It's kept up to date with the Windows version. Same version number!.

Yes, you can choose to only use the open stack (just a newer version of your current, with the necessary MESA packages etc included), you'll then just get a version newer than even rolling release distros. Only can be a little hassle to roll back to the one the comes with Ubuntu if you don't use the uninstall script correctly (once I was left without a working graphics stack, but the apt log shows which packages is replaced, and 1 did not get reinstalled for some reason). And yeah it's supported on RHEL, CentOS and kind of on openSUSE because of that (it's mentioned in the docs, which are linked to from the release notes). The AUR one seemed like a hassle because of this being specific for cards, but I only checked like summer/fall 2019. The driver is specific to your card/series too.

Maybe they're waiting for the next generation of Ryzen? :man_shrugging:

GPU vertical mounting trend

Interesting. Not sure there's space now with my monster Noctua, but it was certainly possible with the 3600X + stock fan.

BUT no reason here, take a lot at this:

$ sensors

amdgpu-pci-0900
Adapter: PCI adapter
vddgfx:       +0.72 V  
fan1:        1003 RPM  (min =    0 RPM, max = 4500 RPM)
edge:         +31.0°C  (crit = +94.0°C, hyst = -273.1°C)
power1:       30.02 W  (cap = 120.00 W)

That's just 2°C above Ambient. I've frequently seen it at Ambient Temperature too!.

Running Unigine Superposition (free download), it maxes out at 72°C, well below the 94°C critical. Windows maxes out at 73°C.

Main reason to install the driver from amd.com before was to get some more data than sensors showed before, but suddenly with a kernel and/or graphics stack update the one since like fall 2019 has included all of the data.

Maybe I'll try the latest Pro driver again to see if it gives better performance. Why not? It just installs a local apt repository, meaning the driver will not update unless you do it manually, which how it should be. If you manually install GPU drivers on Windows instead of relying on Windows Update, it never updates your driver either. There may always be regressions. Often a "Check for updates every X days" is on by default, BUT:

I've disabled such for all my very few 3rd party apps on Windows so it never does anything when idle, not even show a notification (requires service). Some I let check when opened, but IIRC none set to that. On macOS that's what I choose. Apart from the benchmark programs Blender Benchmark, Geekbench and Unigine Superposition, I only have these programs installed:

Corsair Toolbox
Does not run on start-up, not any service - a manual check for updates to it AND firmware updates is actually needed and confuses many dumb users as at least for new HW often an update to the tool is needed before the firmware update appears. Was the case for my 2 MP600s
Corsair iCue
To control my RAM chips lightning, which is apparently stored in the chips, as the settings are kept on Linux. The service it activates is disabled as it does nothing really. Just gotta start it before opening the program if I want to make changes. Maybe by default the program itself runs on start-up as a tray icon, but if so that's totally disabled, so it's totally disabled
Bitvise SSH
The recommended by previously popular PuTTY, as it has SFTP browser and everything. Does NOT start on start-up. Closes tray icon when exiting last terminal. Nice to have the icon visible in case you forgot a terminal is open so...
Sublime Text 3
Kind of need a proper editor on Windows. Comes with 2 Notepads now! At least 1 finally supports Linux newlines and then supports JSON files, but still they're useless for anything other than the basics so Sublime is set to open all text files
VLC
Can only check for updates at start-up but I've disabled that too. It's just there in case I need to watch a video on Linux
Windows Search Indexer
If you rely on searching for files, disabling it is a bad idea. But I know where all my stuff is. I do let it Index the Start Menu, but no folders. It only seems to activate then when a new program is installed/uninstalled. But the service can be disabled
Windows Disk Trimming
Set to be done manually. Only that dedicated disk for Windows with little activity. My partition on the NVMe has nothing on it so don't need trimming anyways. If any HDD has a Windows file system Defragmentation will be in the same place (now called "Optimize") for disks. On Linux I let systemd-fstrim.timer enabled as default, about once a week I think, at night or whatever
Windows Update
Set to "Pause for 7 days". Very important to check that it's paused beforr benchmarking. I just let it check to check if haven't booted for long, trigger it manually. In case there's a new build and security updates. No "Optional updates" installed. It wants to install something for the below, but I know it breaks it, so no thanks. All the above Windows services are particularly important for an idle system
Logitech SetPoint
Needed to map a special key on my Performance MX to something else than default, which IIRC is Zoom. I love having that button mapped to the Home keyboard button. Very practical on websites to go to the top, and certain software brings you... Home. That must be running, as a tray icon, but once it has started and mapped the key it does nothing, zero usage seen from it. Mind you I have the same setup on Linux with xbindkeys starting at desktop login. I've also set the big thumb finger button to open the GNOME list of programs like the Super button does:

$ cat .xbindkeysrc                                                  
"sleep 0.25 && xte 'key Super_L'"
b:10

"xte 'key Home'"
b:13

olejon on 28 Jun 2020

Didn't you say something earlier about having made CoreFreq "compatible with k10temp"?

It shows the High value for the Ryzen 3 CPUs and code for it could be used to trigger the HOT indicator?

olejon on 30 Jun 2020

Didn't you say something earlier about having made CoreFreq "compatible with k10temp"?

Yes if those k10temp Kernel prerequisites are detected and you build _CoreFreq_ with this argument:

make HWM_CHIPSET=COMPATIBLE

...
corefreqk.c: In function ‘Core_AMD_Family_17h_Temp’:
corefreqk.c:7988:5: note: ‘#pragma message: "Building with Kernel amd_smn_read()"’
...

The pragma message above confirms that both driver use the same Kernel API (which protects SMU accesses by a mutex).

It shows the High value for the Ryzen 3 CPUs and code for it could be used to trigger the HOT indicator?

This is a hard coded value if I don't make a mistake. See its source code.
As I told you I'm trying to avoid as much as possible hard coded values.
I've read some docs that SMU has some sticky bits when thresholds exceeds: tbc.

cyring on 30 Jun 2020

Referring to issue #191 , can you please give a try to the develop branch with any of your hypervisors.

What I found in KVM and WSL Linux drivers source code is that they provide a _virtual_ counter for the CPU usage.
This counter is read through a MSR register.
Behide the hood, it appears that this counter is relying on the Kernel computed runtime usage per CPU.
Not the usual hardware counters (APERF/MPERF or Intel PMU fixed counters: they crash immediately)

The results are quiet strange:

The max frequency ratio on a single vCPU is one bin less than the reported value by the real processor
the computed idle state in C1 is a few percent above than bare-metal case

Don't ask for all the other sensors. Those registers are not virtualized: temperature, power, voltage, etc
But we get the virtualized ISA & CPUID features capabilities and also the VM tasks scheduling.

Thanks for any help.
There are some much Virtualization cases that I'm not sure _CoreFreq_ driver is strong enough to cover them.

cyring on 30 Jun 2020

Thanks for that make option. What exactly does it do?

Yes I'm having a hard time finding that 70°C is some AMD spec. 95°C is though. Some people report seeing up to 105°C in Prime95 and such (BUT people say that tool if AVX is enabled is not made for these CPUs), and that it simply would shut off (not destroying the CPU). Others say: "No, it should not but simply downclock". Maybe because it got so hot so fast?

Many reports 80-88°C in stresstests, which I can only get with PBO = Enabled in Tweaker, not AMD's PBO, which they seem to call AutoOC, although that option does not exist in UEFI they have to refer to Ryzen Master's Auto Overclocking Mode. This article explains stuff pretty well you'll probably understand. About PPT, TDC, EDC... which is limited to the Motherboard's limits if enabled.

What's confusing is that ASUS calls it Precision Boost Overdrive both places (under AMD Overclocking and Tweaker) BOTH places in the UEFI. With the exact same Help String, but the effect is very different. The AMD one does little. The Tweaker one does increase performance, especially with All Cores stressed.

With PBO disabled we logged 11.6A multithreaded and 3.3A single threaded (139.2W/39.6W), and with PBO and AutoOC maxed we logged 14.2A multithreaded and 3.4A single threaded (170.4W/40.8W)

... and they don't see too concerned about that 170.4W reading. Also they say:

AMD intends PBO and AutoOC to be used in conjunction with each other, with the AutoOC offset benefitting single-core boost and PBO benefitting multi-core loads that bump up against power limits

... I don't really see that. I get the 4.7 GHz on single core without ANY AMD Overclocking.

Also if I understand the 3950X correctly it's supposed to manage about 1.45 V. If I enable PBO in Tweaker I get 1.28 V while I DO get the 200 MHz bump on All Cores, which must be why I'm seeing 193-203 Watts then. Otherwise never above the spec of 142 Watts.

Question: With this in mind (Temperature = OK, voltage = OK), are you still worried about the Watts?.

They use a Motherboard that, with this enabled, sets the PPT all the way to 1000 Watts BUT according to them that's just the Motherboard/BIOS and the manufacturer can very well change that to even more in a BIOS update. It SHOULDN'T matter, and for me it doesn't as it sets PPT = 395 Watts BUT never reaches that, it's just hypothetical, so CPU limits it?.

Question: People also say Manual OC is superior to any Built-In OC setting, like a drop-down. Now there are SO MANY different configs people post for their UEFI on forums. ONE THAT SEEMS TO REPEAT ITSELF = Voltage Offset DOWN with -0.05 V, for better performance. If say setting it to -0.1 V it just performs worse. What exactly does it mean to set the Voltage Offset DOWN?

PS: I set Linux governor to performance on ALL policies in sysfs (ACPI drivers) and it does the same as on Windows. All Cores ALWAYS runs at Base Clock = ~3.5 GHz. It increases CoreFreq minimum MHz with approx. 20 MHz up from 0002 to 0022. But it does perform a little better with Conic Compute. That extra MHz seen with Conic Compute is not because of the overhead, right?.

Please read. This is a VERY popular YouTube channel: https://www.gamersnexus.net/guides/3491-explaining-precision-boost-overdrive-benchmarks-auto-oc

olejon on 1 Jul 2020

Many questions above, I don't master the architecture enough to give a precise answer beside my observations.

Yes, my BIOS is set to lower the temperature with:

CPU voltage minus offset down to 1.4V UEFI and measured idle 0.9V CoreFreq
PBO disabled, CPB enabled
NB C-States enabled
DOCP for DDR @ 3600 MHz
IOMMU & AMD-V enabled
Remaining opts in AUTO

While coding, temps are 32°C
Conics limited at 70°C
Single Core frequency can max at 4.7 GHz

CoreFreq build option HWM_CHIPSET=COMPATIBLE allows to run concurrently _CoreFreq_ and lm_sensor/k10temp

cyring on 1 Jul 2020

Then we have the same UEFI settings except I haven't touched voltage (I do things like disable the serial port + legacy USB + SMART TEST for HDDs (like why...), change Boot options, disable MB RGB, disable F1 if error at boot because server services, enable "Last state" if Power Loss" because server services and tune Fans Duty Cycles and Temperature Thresholds, manually not the stupid curve which you use the Mouse to edit, although I see overclockers using that, just changes those values)
Mine also idles at 0.91V as I don't have any OC on except D.C.O.P. like you
C-states are on, but IDK if you refer to the CoreFreq driver there(?)
With AMD-V do you mean SVM? It's the option in my UEFI that says "Enable CPU Virtualization"
IOMMU I haven't touched, what does it do exactly?
PS: I can run sensors, updated every X milliseconds, alongside CoreFreq without compiling that way, no problem. Both show the same temperature so still don't get what it does(?)

Found some Windows discoveries yesterday:

Use the amd.com chipset drivers, not the ones in your Motherboard's website. They're outdated. Can be uninstalled and those from amd.com installed without even rebooting
Found that Ryzen Master's "Check for update" doesn't work. Downloaded it again. Uninstalled previous first, which removed the Power Plans (useless but luckily they're just in that package), and installed the downloaded. Immediately saw it was newer, because new Splash Screen. Some more values at top. Otherwise not much
Windows is at least now stable at 4m27s (+/- 1 second) in Blender classroom and Linux 4m (+/- 1 second)
Get 4.7 GHz one 1 Core but really just see the 4700 MHz+ if system has been idle for a while. Have no idea why. Otherwise about 4985 MHz which well, is basically 4.7 GHz
On all cores, about 3850 MHz after a little while with Conic Compute. Bad? Boost of 0.35 GHz... Windows gets higher but doesn't win any tests
Can't find an easy way to stress just 1 Thread in Windows. All tools stress 2. Although CPU-Z has 32 listed and I choose 1, it stresses 2
Then downloaded the official ROG CPU-Z version (choose the one from CPUID's website, not the other one (Freeware site, dunno why there's 2 links) and saw there that there's no more choosing 1. There's 2, 4, 6 etc
Anyways just choosing "Stress CPU" gets better MHz than Conic Compute but not Atomic Burn or CRC32 Compute really, BUT CPU-Z stressed the CPU more than those, seen by Temperature and Watts, so...
You might want to check out https://rog.asus.com/downloads/
The RealBench Tool might be interesting for you. It's all FOSS in just a ZIP file and you don't install anything. Runs like GIMP etc. Mine crashes (probably because no ROG MB?) at H.264 (or GPU?) and the "Heavy Multitasking" (think HandBrake test there fails, there's a log window with progress), but the Image one using GIMP and the one using LuxMark work

Installed the AMD Pro driver yesterday on Linux. So far so good
Gives
- Pro OpenGL
- Pro Vulkan
- Pro OpenCL (apart from uninstall/revert script really just a practical binary in /opt with more than the standard clinfo command
Now possible to benchmark OpenCL in Blender like on Windows
Sorry for having said radeon driver before, it's of course called amdgpu now. There's a cool GUI GitHub project I wanna check out (has AUR package, otherwise compile and systemd daemon if wanna run as regular user). So far only played with those values in sysfs, like Fans PWM values, but there's plenty of GPU OC settings AMD patched into the kernel in 2019
Same here, depending on ambient, can be 32°C and max 70°C if stressed. Wonder how it would've been if there was space for the 2nd Noctua Fan. Thought about changing the chassis fan on the back to the Noctua, more silent and possibly better, basically straight behind the heatsink, but discovered the Fan Head Connector is not not possible to reach with this heatsink :disappointed:

olejon on 1 Jul 2020

Then we have the same UEFI settings except I haven't touched voltage (I do things like disable the serial port + legacy USB + SMART TEST for HDDs (like why...), change Boot options, disable MB RGB, disable F1 if error at boot because server services, enable "Last state" if Power Loss" because server services and tune Fans Duty Cycles and Temperature Thresholds, manually not the stupid curve which you use the Mouse to edit, although I see overclockers using that, just changes those values)

Almost the same: SATA disabled but I keep Q-Code leds which display the CPU temperature on MB !

C-states are on, but IDK if you refer to the CoreFreq driver there(?)

That's the all point. What are those BIOS NB C-States !

With AMD-V do you mean SVM? It's the option in my UEFI that says "Enable CPU Virtualization"

Yes, SVM

IOMMU I haven't touched, what does it do exactly?

Virtualization of devices. Such as GPU PCI pathrough

This 3950X can virtualize an EPYC 255 VCPUs with non pain

CoreFreq_KVM_EPYC

PS: I can run sensors, updated every X milliseconds, _alongside CoreFreq without compiling that way, no problem_. Both show the same temperature so still don't get what it does(?)

Without the compatible Kernel API, the SMU may be flood by both drivers.
It takes a variable amount of time until the concurrent access issue appears.

... it's of course called amdgpu now. There's a cool GUI GitHub project I wanna check out (has AUR package, otherwise compile and systemd daemon if wanna run as regular user).

I did some readings about the MESA stack for Radeon 5000 series: it looks well supported according to devices database. I will however wait Xmas for the next RDNA2 GPUs

... but discovered the Fan Head Connector is not not possible to reach with this heatsink disappointed
You bet ! I also opened a ticket at Fractal to have a fan cables topology... 2 weeks without any answer.

cyring on 1 Jul 2020

👍1

Ah, well could've read Wikipedia about IOMMU. The UEFI Help String kind of makes sense then. Maybe it's even in the manual (much UEFI stuff isn't).

Is IOMMU needed for that CPU virtualization or was that just a cool demo? Haven't had problems emulating other CPUs with IOMMU disabled. Guess I'll leave it disabled unless... There's a reason for it to be enabled just running VMs (on this we're talking desktop VMs, on another headless host I have my headless VMs)

The Linus Tech Tips video where they make a Hackintosh (but fully working) using Manjaro + KVM + QEMU, and 2 GPUs, one dedicated to the macOS VM, that's a cool one. It's technically illegal but they show the whole process maybe except a part of after having downloaded macOS to a file (on a Mac which should be legal), how to make it a bootable disk for the VM (but they say something like "see the forum, user X explains it"). They have to use various patches from GitHub to QEMU and some special things in the virsh edit XML file, but IIRC the GPU they just choose from the drop-down of virt-manager. Intel of course, components known to be supported by Apple.

Whole point was to make a PC as good as the upcoming new Mac Pro. Only thing they couldn't match was the amount of RAM the model that costs the most can have. But enough slots for a lot. The Xeon they used was like 1 version older as the one to come with it, and announced, wasn't released yet.

CPU passthrough mode for sure, works fine here, then a VM shows my exact CPU.

Cool thing was: It outperformed the best iMac Pro, and those aren't bad, and they are expensive. In a Linux VM. One monitor for Linux and the other for macOS. They got iMessage and FaceTime working which as far as I know has been a thing Apple has locked Hackintosh users out of. Some patches or whatever. Long video so they show the GitHub stuff in like 16 x speed. Can always slow down with YouTube. Later they made it work on bare metal. And compared. Did better but not that much, really. Been a long time since I've seen those channels now. It was an ongoing project.

Wonder now, if they have managed a fully working "Mac Pro" (without being a Mac per se), was supposed to be fitted in an old Mac Pro chassis ("the good old one" that was a normal PC basically) that is cheaper and faster, with native macOS. But IIRC the Mac Pro, and iMac Pro, after checking a comparable HW PC build, aren't really that expensive. Can even be cheaper, if you mess up. Meaning little profit for Apple on those (but could they take even more and get customers?)

On Intel I can use host-model, but the file containing lists if CPUs didn't include Ryzen 3 last I checked, although long ago now, many QEMU updates ago, but with CPU in host-passthrough I got my 3600X recognized as that, otherwise "AMD EPYC".

My Windows 2000 shows AMD EPYC as I haven't bothered with tuning it. Just wanted to play with it. Need to change from just matching topology to passthrough and see if Windows 2000 says 3950X!

olejon on 1 Jul 2020

And duh, I should've put on those fans header extender cables that comes with both the chassis (and the Noctua) to that fan on the back before putting on the heatsink. Too trigger-happy... Was watching out it didn't touch anything. Forgot access. Also 1 NVMe is below, the CPU bound. But very satisfied so no problem...

olejon on 1 Jul 2020

Can you subscribe to an Apple developer account within Hackintosh ? Will you be able to push apps to the store ?
If not, you won't get any Apple developments support.
Most of the API kits are closed sources...
You can see IOMMU as a virtualization option, but:

VM shows my exact CPU

is not about path-through, but a Hypervisor catch and translate

cyring on 1 Jul 2020

Don't know. Community is big so maybe ask some project in GitHub that only host some patches. Surely developer knows. If Apple ID and everything works, it thinks it's a Mac... But no guarantee. I have an account, think I created it for some more advanced forum, or iOS beta, always get invited to WWDC
But not few years ago one could develop fine on Windows/Linux, and emulate on X physical devices, but signing and publishing had to be done on macOS
Aha. I see some better performance with it on the Intel headless. But go for the host-model with exact and no fallback, because; only use the actual cores not HT. That CPU has a security bug with HT that apparently can't be fixed (lives in the CPU).

olejon on 1 Jul 2020

For some Zen processors, the IOMMU state can now be supplied by _CoreFreq_
Can you give a try to issue #193

cyring on 1 Jul 2020

Oh that I can confirm immediately. I checked it. See below points too

Screenshot_20200701-160917__01

I've adjusted my cfrun script, which if run with --update downloads the latest code from main branch to compile like this: make MSR_CORE_PERF_UCC=MSR_IA32_APERF MSR_CORE_PERF_URC=MSR_IA32_MPERF HWM_CHIPSET=COMPATIBLE
Q: Since NMI Watchdog is on, is the APERF/MPERF needed on AMD? Any disadvantages at all?
Q: Does it hurt to compile with the same command on Intel? Regarding HWM (HWMonitor?). Since I want to keep my script the same on all PCs...
This was cryptic (probably even for OC enthusiasts and pro sysadmins):

Without the compatible Kernel API, the SMU may be flood by both drivers. It takes a variable amount of time until the concurrent access issue appears

First of all I always unload CoreFreq completely after use, just a cfstop command. BASH alias to main script, and that script is added to the NOPASSWD list for sudo, like other tools as iotop
Q: Is an "SMU flood" then a problem as long as k10temp is always loaded, while corefreqk only when I use the UI? What signs would an "SMU flood" show? What negative impact does it cause?
The k10temp module will load on all AMD compatible systems just like coretemp on Intel. So, shouldn't it by default be compiled like that? At least on AMD if that's the only arch needing this? The README.md doesn't even mention this compile option, and make help says NOTHING about what it does. In fact the README.md is confusing because at first it says to compile with fixed counters, then in the instructions below that actually shows the complete procedure of cloning the code etc, it's enough with simply make? No mention of APERF/MPERF there either, just says to turn off NMI Watchdog in the kernel parameters (not everyone wants to do such things), but you said isn't necessary the way I compile with MPERF/APERF, right? And make help explains nothing about those cryptic options
So maybe make README.md a little clearer. It's another thing with make help, kind of, but you expect actual Help, no? Just a tip to avoid questions like these :smiley:

Be aware: Just opening Ryzen Master with Default OC Mode sets this, without asking or anything:

Maybe only if DCOP is found enabled, but everyone does that of course. Not that it does any harm... EXCEPT: Certain UEFI Settings under AMD Overclocking can NOT be reset by a CMOS reset and you can brick your motherboard (confirmed on a forum for ALL Ryzen motherboards)
NOT that this particular Setting will brick anything as all values below are Auto, so
But I saw this then clearing the CMOS with a screwdriver, that afterwards, some AMD Overclocking Settings were still set, they're stored somewhere else... Careful in that section in other words. Tweaker is then safer...
Got my Windows 2000 VM to show "AMD Ryzen 3950X 16 Core", yes :stuck_out_tongue:

Yet to try the develop branch. What do you have in mind? Try say Manjaro in KVM + QEMU with passthrough and see how CoreFreq works? I'm gonna install Manjaro XFCE anyways... Others in mind like VirtualBox? Even on Windows? Or even the Ubuntu Windows subsystem?
Yet to make a .deb. Will convert my script to use systemd first. Better for my way to load/unload CoreFreq (that includes the daemon) anyways. Per now it's basic BASH
Q: Any possible disadvantages of having corefreqk and/or the daemon loaded all the time? Just curious. Don't like to have any module not installed by apt or kernel default loaded all the time, simple as that. And the daemon starts processes to... Maybe I'm paranoid...
Suggestion: Please-Please move ~/.corefreq.cfg to ~/.config/, as per Freedesktop standards. Even pure proprietary CLI software you wouldn't expect to do so, does that now, and best if it's in its own folder, that way, no file extension needed, say ~/.config/corefreq/config
And the other, add (W) and (J) behind the values at the bottom at the W view... It's just not logical that it's only behind the last one there, for say a new user

olejon on 1 Jul 2020

NMI_Watchdog, MSR_CORE_PERF_UCC, MSR_CORE_PERF_URC, all of them concern only Intel processor.
Building for AMD, there are no other registers choice than APERF/MPERF with which NMI_WATCHDOG has no impact.

True, documentation needs improvements.

.corefreq.cfg is supposed to be XDG compliant but code is buggy. I also want it to be in .config directory and home as a fallback location: sign that XDG environment variable is not detected.

cyring on 1 Jul 2020

Thanks! Did I understand you correctly that compiling with APERF/MPERF doesn't do ANYTHING on AMD then and I can keep it in my build function in the script?

And HWM, on Intel is it OK to compile it with COMPATIBLE although it's Intel?

olejon on 1 Jul 2020

Intel: you can choose among MSR_IA32_APERF, MSR_IA32_MPERF, MSR_CORE_PERF_FIXED_CTR1, MSR_CORE_PERF_FIXED_CTR2,
and MSR_PPERF
AMD: MSR_IA32_APERF and MSR_IA32_MPERF

cyring on 1 Jul 2020

HWM_CHIPSET can be equalled to W83627 (SuperIO for various sensors, time of Intel X58) or COMPATIBLE (for AMD SMU within Northbridge)

As default: _CoreFreq_ accesses SMU on its own.
COMPATIBLE build directive will make my driver goes to the SMU through a protected Kernel function.

cyring on 1 Jul 2020

Ok, so I keep APERF/MPERF on both Intel and AMD without problems? Yes/No
APERF/MPERF basically required on AMD? Yes/No
So COMPATIBLE only for AMD? Yes/No

Still unsure: Which of those counters does simply running make use? Is it smart and detects? Seems like compiling with COMPATIBLE is recommended on AMD but not default, if I understand you correctly?

olejon on 2 Jul 2020

Ok, so I keep APERF/MPERF on both Intel and AMD without problems? Yes/No

YES

APERF/MPERF basically required on AMD? Yes/No

YES

So COMPATIBLE only for AMD? Yes/No

YES

Still unsure: Which of those counters does simply running make use? Is it smart and detects?

YES, a plain make will _find_ all definitions to build _CoreFreq_ but not the processor architecture.

Seems like compiling with COMPATIBLE is recommended on AMD but not default, if I understand you correctly?

True
( without COMPATIBLE, the default, the instructions path to reach the SMU is shorten. )

cyring on 2 Jul 2020

Fyi: Among the last develop changes, the unit is now added to the power domains.
CoreFreq_UI_20200702

cyring on 2 Jul 2020

Thanks! Then I will remove COMPATIBLE from my Intel PCs script and rebuild, although it's unclear still if it may cause problems on Intel (don't see any)? Yes/No
If APERF/MPERF is required to add to make for AMD I suggest you Highlight that in README.md, or you make make a little smarter and it does it "automagically"
Cool. That's a more clear UI, like other tools, always show measurement unit
This because on AMD CoreFreq was the only tool to see temperatures until k10temp added Tdie: X°C and Tctl: X°C

* Both those show the same, and the same as in CoreFreq, although it has 1 decimal and CoreFreq *rounds that value Up
Still I think CoreFreq can become a go-to tool for many on both Intel and AMD if instructions are clear for regular users! Especially monitoring but stresstests are good, HOT indicators are good, and tweaking surely someone will use!
Documentation is crucial. A clear README.md with links to the Wiki instead of a long list of stuff. Say one link to installation (divided into Intel and AMD sections) and one link for FAQ. Regular users don't understand things like registers and counters. This might be under "Advanced Tweaking" documentation, for example
Use the GitHub pages to create a simple website for it so you get github.io/corefreq? To promote it better, with screenshots of latest version on Intel and AMD. Very easy, just creates 1 HTML + CSS, in a standard branch so can be edited just on GitHub without cloning and pushing for small tweaks
Are you enough satisfied with it now for a Debian/Ubuntu package? Yes/No
Or still consider it too experimental? Yes/No

Imagine if someone uses the JSON to implement it into GNOME System Monitor and others, for instance... Windows Task Manager has gotten pretty good if you use the Performance tab it has Usage (in MHz not just a graph), Temperatures and Watts for CPU/GPU, and more readings. Adding GPU and more CPU readings like temperature and Watts lack (only that for a monitor).

Add your PayPal email to the About View. Just a humble thing, so you can buy more HW, and a text like "Thank you. Donations will be used to improve CoreFreq and ensure compatibility with New Hardware". Nothing more... I'd happily donate. Can I? Which email?.

olejon on 2 Jul 2020

👍1

Suggestion: Please-Please move ~/.corefreq.cfg to ~/.config/, as per Freedesktop standards. Even pure proprietary CLI software you wouldn't expect to do so, _does that now_, and best if it's in its own folder, that way, no file extension needed, say ~/.config/corefreq/config

export XDG_CONFIG_HOME=$HOME/.config
corefreq-cli

[CTRL]+[x]

cat $XDG_CONFIG_HOME/corefreq.cfg
61,14,7
70,3,5

Don't you get the same location ?

cyring on 3 Jul 2020

[WiP] Some changes in the Power footer:

cyring on 5 Jul 2020

Nice. I see this is in the develop branch, so I'll wait
Kind of don't want to export XDG_CONFIG_HOME=$HOME/.config every time. I mean even htop uses ~/.config/htop/htoprc and has for a long-long time (no XDG needed)

Just checked on VM as root (sudo -i as su) without any XDG set, then, running env:

In fact if the directory does not exist, htop creates ~/.config/htop/htoprc (as if mkdir -p)
Don't know how htop does it, but see source code, or do it your own way, detect at least if directory exists, or even create as htop if does not exist.

Only see these are set logged into desktop, so:

$ env|grep -i xdg
XDG_MENU_PREFIX=gnome-
XDG_VTNR=2
XDG_SESSION_ID=64
XDG_SESSION_TYPE=x11
XDG_DATA_DIRS=/usr/share/gnome:/usr/local/share/:/usr/share/
XDG_SESSION_DESKTOP=gnome
XDG_CURRENT_DESKTOP=GNOME
XDG_SEAT=seat0
XDG_RUNTIME_DIR=/run/user/1000
XDG_CONFIG_DIRS=/etc/xdg/xdg-gnome:/etc/xdg

olejon on 5 Jul 2020

Nice. I see this is in the develop branch, so I'll wait

Thanks.
yes, putting other UI things before commit ...

Kind of don't want to export XDG_CONFIG_HOME=$HOME/.config every time. I mean even htop uses ~/.config/htop/htoprc and has for a long-long time (no XDG needed)

Just checked on VM as root (sudo -i as su) without any XDG set, then, running env:

In fact if the directory does not exist, htop creates ~/.config/htop/htoprc (as if mkdir -p)

Don't know how htop does it, but see source code, or do it your own way, _detect at least if directory exists_, or even create as htop if does not exist.

Only see these are set logged into desktop, so:
$ env|grep -i xdg
XDG_MENU_PREFIX=gnome-
XDG_VTNR=2
XDG_SESSION_ID=64
XDG_SESSION_TYPE=x11
XDG_DATA_DIRS=/usr/share/gnome:/usr/local/share/:/usr/share/
XDG_SESSION_DESKTOP=gnome
XDG_CURRENT_DESKTOP=GNOME
XDG_SEAT=seat0
XDG_RUNTIME_DIR=/run/user/1000
XDG_CONFIG_DIRS=/etc/xdg/xdg-gnome:/etc/xdg

At the time, I have to decide a saving strategy and I found common among major software to follow XDG; even if, most of them use workarounds when XDG is insufficiently configured by distribution.
Chromium is part of them (for instance)

cyring on 5 Jul 2020

develop released
The new label CCX means that the C-States are enabled.
Not sure if changing C-State in BIOS is well reported by _CoreFreq_ ?

cyring on 5 Jul 2020

Hello,
Could you please give a try to the latest develop commit.

https://github.com/cyring/CoreFreq/issues/195#issuecomment-654290433

cyring on 6 Jul 2020

Thanks for Watts now showing!

Another suggestion if possible:

It is very practical to be able to stress a specific Core, to check best cores performances
But CoreFreq can only stress 1 or all
Is it possible to add an option to Tools > Turbo Select CPUs... (notice the plural)
Like X (check) Cores to stress. Stressing 2 of the best on same CCX should give better than stress -c 2 as then Linux choses and they may change on the fly

* My PC must be on 24/7 now without time for any system "locks" as you asked me to test... No freezing thanks. It has never happened in fact

Haven't bothered to search for such a tool on Windows, little annoying that CPU-Z doesn't stress 1 Thread when told so, BUT latest changelog says "Added support for Ryzen 3900X" so seems it lacks 3950X still. Surely there is a tool
Another thing noticed, then using Single Core stress, say CPU-Z set to 1 Thread which stresses 2, the AMD Chipset drivers seem to be intelligent and choose the 2 best Cores on same CCX! (I strongly suggest to download from AMD directly, not Windows update nor your MB's website, works on any X570 MB, Direct Link to download site, not download file)
Installing these drivers fixed the "problem" of 4 "Unknown devices" listed in Device manager in Windows. Not ASUS nor Windows Update ones fixed that. Though I have no idea what they were, didn't take screenshot before/after, forgot it

XDG is insufficiently configured by distribution

Which ones? Ubuntu has it perfectly. Say changing language would change content of ~/.config/[user-dirs.dirs|user-dirs.locale] Google Chrome certainly does it correctly, and not much is different from Chromium on Linux apart from DRM playback. Was a hassle to fix this on Arch, find right package(s), heck even getting~/.config/autostartto run.desktop` files. Say Dropbox relies on that.

Config to ~/.config/google-chrome/
Cache to ~/.cache/google-chrome/
Any app installed, and chosen to open as window (right click), or website added (like Twitter is ideal for this for me) added as Menu > More tools > Create shortcut, .desktop files + icons added to ~/.local/share/[applications|icons]. Also added to Desktop folder by these in case such a DE, but can be deleted. And if GNOME, ~/.gnome/apps (not sure what that's about but think it has to do with integration into Notifications etc, maybe being fased out, as they're identical, or for backwards compat.)
Only big "sinner" is Firefox, probably for legacy reasons, puts config in /.mozilla, but cache correctly in ~/.config/mozilla/firefox/. SURELY legacy/backwards compat. reasons
Several CLI-tools, including proprietary use these, heck even speedtest.net's official CLI tool. A few recently installed program that do too (talking about the 3 directories): Wireshark, Skype, Blender Benchmark, Unigine Superposition

olejon on 7 Jul 2020

Yes, adding CPU to a running stress is definitely something I want to do.
Impacts on the Daemon threads scheduler, I'll go carefully on these changes...
Since I built my Ryzen setup, I've tons of tiny features to develop: most of them can be a few hundreds of C &
asm instructions which also required specs research.
I also have to test for non regressions on any processor I can access to. For example I noticed that K8 is not enough covered...
I'm lacking Bulldozer: I still don't know how _CoreFreq_ works with them... Lot of server MP processors need to be completed...

In answer to your previous post about donations , what I need in fact is man power: a team of system programmers to help with other architectures and their specific features...

cyring on 8 Jul 2020

So it's possible to add "Stress 1 and 2" only?

Hmm, ok. That will be hard I guess. Maybe some subreddit has people with the knowledge? I've seen post deep into stuff. Also you know about projects also reading stuff CoreFreq does.

I can donate for that too, meaning if they want some money to help you. And as said new HW. Let's say you had no Ryzen 3rd Gen CPU. I would've sent you enough by PayPal to buy the 3600 (not X-version), which basically works the same. OR I would have sent you my 3600X :package:

... although nice to have a spare CPU. More I need a spare GPU. Since no iGPU, system won't work without one, at all. Can always go to a physical store the same day, but... won't necessarily have the card I want. Guess they have more the top of the shelf ones.

olejon on 8 Jul 2020

So it's possible to add "Stress 1 and 2" only?

Yes possible

Hmm, ok. That will be hard I guess. Maybe some subreddit has people with the knowledge? I've seen post deep into stuff. Also you know about projects also reading stuff CoreFreq does.

I can donate for that too, meaning if they want some money to help you. And as said new HW. Let's say you had no Ryzen 3rd Gen CPU. I would've sent you enough by PayPal to buy the 3600 (not X-version), which basically works the same. OR I would have sent you my 3600X

... although nice to have a spare CPU. More I need a spare GPU. Since no iGPU, system won't work without one, at all. Can always go to a physical store the same day, but... won't necessarily have the card I want. Guess they have more the top of the shelf ones.

For the rest, don't worry, I'll handle. Up to now I've been able to cover hardware with just their specifications.
It will just go slowly...

cyring on 8 Jul 2020

Yes possible

Great, really looking forward to it! Would make it a much more powerful tool than standard Windows ones, although I'm 100% sure there is software for Windows for this
Partly because it seems like my :star: on my CCX 2 performs worse than the "Grey Dot" on my CCX 1, and "Grey Dot" pretty much equal to "Grey Star" on CCX 1. Having THAT functionality in CoreFreq would allow me to confirm, meaning stressing say 2 Cores and see the Frequency the CPU manages then compared to 2 other selections. Of course this would be great on Intel as well
Of course the Best Cores I have to find in Ryzen Master but those don't change so it's just to write it down or screenshot it
As said, seems like Windows handles Single Core(s) stress better than Linux, using those exact mentioned Cores (Good Star + Grey Dot). Probably the mentioned AMD Chipset drivers
BTW on all forums people say you definitely want the latest of those, they don't even require a reboot and the installer is dead simple but nice, and also BIOS for latest AGESA. I'm waiting for a new one, and will try if it doesn't say it's impossible to revert, which ASUS writes in the Changelog if that's the case. Still on April one although maybe the new kernel updates have fixed the "BIOS bug" messages I got, as the June one basically only lists an update to exactly AGESA, which I had to Google Search to find WTH it is. It's on Wikipedia even. Did partly understand it
I've found, on AMD, using stress -c 32 to be better than any Conic Compute. It stresses the CPU as hard as seen by Temperature it reaches, which is the main point, and Frequency stays higher meaning higher Turbo % as well (rounded to 113% instead of 111-112%). Say Atomic Burn/CRC32 of course gets even higher Frequency and then Turbo %, but really no stress, nowhere close on Temperature
This was NOT the case on Intel. Is Conic optimized for Intel? Would love to see another or a refined Conic that's harder on AMD
PS: Forget one UEFI Setting in my latest Benchmark. There's Zero reason to set AMD OC PBO Scalar = 10X. Let it stay at 1X. Default is 2X. Saw a post about it. Decided to check and indeed no difference. It might even be bad but that seemed more speculation by the poster. Most OC testing web pages about PBO simply set it to 10X. I don't think it hurts. The CPU handles AMD OC just fine, because remember, these settings are AMD OC in UEFI, NOT Tweaker

olejon on 9 Jul 2020

Yes possible

Great, really looking forward to it! ...

Work in Progress ;-)

cyring on 9 Jul 2020

👍1

Yes possible

It is available in the develop branch and applies to the Turbo Select CPU...

CoreFreq_Cumulative_Stress

Basically, when all CPUs are added, you get the same as Atomic Burn

Because Conic algorithm has been made for parallelism, it is not part of this delivery.
_EDIT: neither the others_

Have fun

_EDIT_: If threads get stuck stressing, kill -9 $(pidof corefreqd-ring1) and please let me know.

_EDIT_:

_CoreFreq_ provides same features for both Intel and AMD, except architectural specifications.
Thus, the stress tools apply to all brands.
I don't spend much time on Windows and its software ecosystem. Beside WSL, b/c I had hoped I could port _CoreFreq_ to Windows for free. Lol
I'm a new player in this field of performance applications, so I believe you that other Windows software are stressing better.
However, the less I look at others, the more added value and creativity I hope to bring.
_CoreFreq_ users are also of a great help: ideas and synergy will help to make some unique features.
Notice in the Roadmap issue that AVX is something to do because those instructions have a _huge_ impact on the Processor operating work.

cyring on 10 Jul 2020

Last commit, you're getting the scheduled state of each CPU
This sticky window allows to add CPU when Turbo Select CPU.. is the current stress tool
Only [F10] or STOP can remove CPU(s) from the scheduler
The shared memory is changed: you must fully rebuild and reload _CoreFreq_

cyring on 10 Jul 2020

Awesome :clap: I'll wait until in stable. Just can't afford any system freezes now.

olejon on 11 Jul 2020

With [Turbo Select CPU...] you can now add or remove CPU(s) as will.
This develop version is stable.

cyring on 11 Jul 2020

Great. I'll wait until it's in Stable even so, in case you add other stuff that may freeze the system
Little lazy and use my script, but as said have other things to do
Made this quick systemd service to set the governor to performance at boot (just enable it, and use start/stop as usual). The sleep is needed because if not the service runs before the driver sets it to ondemand. It does exactly the same as using the "Ryzen Performance" Power Plan on Windows. All Cores run at 3.5 GHz all the time. I lose some 15 MHz on Single Core Stress but nothing at All Cores Stress. In fact it improves Geekbench scores somewhat, mainly on Linux!
I've created these Side-by-Side screenshots for new benchmarks. All Blender Benchmark renders at once. The last render is really long and really shows how Linux outperforms Windows. Both totally idle. Not even running a monitor tool...

[Unit]
Description=Governor

[Service]
Type=oneshot
ExecStartPre=/bin/sleep 10
ExecStart=/usr/bin/cpupower frequency-set -g performance
ExecStop=/usr/bin/cpupower frequency-set -g ondemand
RemainAfterExit=yes
TimeoutSec=60

[Install]
WantedBy=multi-user.target

FYI: New version of Ryzen Master. Had a weird simplified UI but clicking Advanced opened the standard. ALSO NEW JULY BIOS, NOT TRIED YET.

olejon on 14 Jul 2020

👀1

FYI: New version of Ryzen Master. Had a weird simplified UI but clicking Advanced opened the standard. ALSO NEW JULY BIOS, NOT TRIED YET.

I did but I don't understand those Vcore:

Windows Balanced Plan
Windows _Eco_ Plan

With Eco, I'm getting a Vcore closed to _CoreFreq_ measurement but Balanced shows Vcore can go as low as 0.4 V . A Vcore value I don't even read in BIOS

EDIT: In French Vcore is "Tension CPU" then you see current value at "Actuels"

cyring on 14 Jul 2020

I've said it before, Voltage on Windows hovers at 0.32 on my system
But jeez choose the Advanced Mode and see the standard window you're known to from before, or "Affichage avancé' for you...
Seriously you don't use your OS in English? I use ALL systems in English... because otherwise they suck, and a good mishmash of languages (and in some instances on mobile devices from Android to iOS you get less features, or they're pushed out later from say Google Play Services)
Use English (GB) if you can't set Region Formats, but that's totally possible on Android and iOS and Linux and Windows (meaning no stupid Fahrenheit and the imperial system), and on Ubuntu it's dead simple through GNOME Control Center, but it really it does it automatically when you set timezone and keyboard layout during install, it gets it, downloading and setting Region Formats to Norwegian ones (basically the metric system, and well some places like in LibreOffice currency sets itself to NOK but that can be changed on right there and then in that software just like autocorrection/spell check)

olejon on 14 Jul 2020

I use both languages: W10Pro license is French. Linux with no locale.
All these reasons which make _CoreFreq_ available in English and French.
Have you noticed that ?
If not, just press capital L
I need translations in German and Spanish. Others not excluded.

Any one in will to translate _CoreFreq_ will have to work on the UI resources file; using the corefreq-cli-rsc-en.h as a template.

What I have not improved yet, is the ascents support.
UTF is the standard, but for low latency purposes, the UI layers are mode of 7-bits characters and color attributes, thus ASCII.
UTF-8 would be a solution but I'm not sure it will be enough to cover the 4 major languages: en, de, fr, sp

cyring on 15 Jul 2020

I can do Spanish. After all I have an international diploma (Institutio Cervantes, Madrid) for most work and national diploma for level needed for being a doctor or lawyer etc. And I've teached Norwegian to Spanish speaking people from all over the world, explained their own grammar to them first as most don't know it, AND of course, got a Spanish GF, talk more in Spanish than any other language, think in Spanish mostly except when having watched a lot of English TV or especially communicated a lot in English like here (though more like really long emails making me think about a follow-up). And of course almost half of my adult life I've lived in Spain and studied a total of 4 years there...
Shouldn't be too much for the CoreFreq UI. Haven't checked the template but I guess it doesn't translate technicalities that shouldn't be
Just don't except anything ASAP
But yeah, it's the 2nd biggest mother tongue on Earth after Mandarin, with English as number 3, while more people talk English of course...

olejon on 15 Jul 2020

Great ! Let me know when you are starting the translation because I have to explain you all the resources constraints you'll face to; such as the string or array length which must be strickly equaled to the English reference statement (but can be blank padded)
True, many anachronisms have been invented by Intel, and it does not make sense to translate them.
Can Spanish be written without accents ? Will it be readable ?
If not then I've to implement first a solution to the accents issue.
Mandarin would also be a great addition. However I have no idea if Linux supports it.
I have been noticed that among _CoreFreq_ users, some are coming from China. I presume they are easily understanding English to come to my software.

cyring on 15 Jul 2020

First of all, NEVER EVER translate the Language List itself. NOBODY does that. Makes it hard to find back to your language of you change it accidentally. I changed to French and suddenly it listed "English" translated to French! Like Anglais IIRC
Normal procedure is either list the languages how their called in that language, like this. Notice the lowercase as only English writes languages in uppercase of these:
- English
- español
- français
Other normal procedure is to show them in English, or system locale but most ignore that because can cause problems for new users, and show them ALL in English, so:
- English
- Spanish
- French
Yes, it's possible to avoid accents, I assume that includes tildes like writing español as espanol for instance. People will get that. Most of the stuff in CoreFreq won't need accents. Some words kind of require it, but many, ma y Spanish speaking people don't use them when texting etc, and many are added only in questions, like Que and Qué have different meanings
Well can't help you with Mandarin other than Google Translate so... I do have a friend, though, maybe. But again, VERY important if you manage to set it to Mandarin, that the Language List is NOT ALSO translated to Mandarin... Hard to find back to say Spanish then!
If it supports Mandarin, then it supports UTF-8 right? Then accents etc shouldn't be a problem, so I guess there is some other constraint?
Remember all accents in Spanish are above letters and don't make them "taller". Like Que/Qué or Energía IS correct and doesn't make it taller while Energia is incorrect but understandable. There are no letters in making it "lower" in Spanish, like ç as in French, and exclamation and interrogation marks like ¡ ¿ are not necessary. Most Spanish speaking people never use them except official texts
OK. I guess I can just translate it into Spanish and you'll tell me if I've broken some constraint and I'll try to fix it? Some stuff is indeed longer in Spanish than English, and vica-versa, but mostly the Spanish words longer

olejon on 16 Jul 2020

Thanks for these advices.
What I will do first is to prepare the resources file and the necessary code to toggle in Spanish.
This will be available in future develop release.

cyring on 16 Jul 2020

I tend to check the ASUS website if there is a BIOS update. Saw there was one from June 17 published. Probably one for you too?

After installing it, Linux boots fine, but see these and journalctl shows the IRQ stuff in Red

juni 24 00:52:21 tux kernel: do_IRQ: 1.55 No irq handler for vector
juni 24 00:52:21 tux kernel:   #2
juni 24 00:52:21 tux kernel: do_IRQ: 2.55 No irq handler for vector
juni 24 00:52:21 tux kernel:   #3
juni 24 00:52:21 tux kernel: do_IRQ: 3.55 No irq handler for vector
juni 24 00:52:21 tux kernel:   #4
juni 24 00:52:21 tux kernel: do_IRQ: 4.55 No irq handler for vector
juni 24 00:52:21 tux kernel:   #5
juni 24 00:52:21 tux kernel: do_IRQ: 5.55 No irq handler for vector
juni 24 00:52:21 tux kernel:   #6
juni 24 00:52:21 tux kernel: do_IRQ: 6.55 No irq handler for vector
juni 24 00:52:21 tux kernel:   #7
juni 24 00:52:21 tux kernel: do_IRQ: 7.55 No irq handler for vector
juni 24 00:52:21 tux kernel:   #8
juni 24 00:52:21 tux kernel: do_IRQ: 8.55 No irq handler for vector
juni 24 00:52:21 tux kernel:   #9
juni 24 00:52:21 tux kernel: do_IRQ: 9.55 No irq handler for vector
juni 24 00:52:21 tux kernel:  #10
juni 24 00:52:21 tux kernel: do_IRQ: 10.55 No irq handler for vector

...

juni 24 00:52:21 tux kernel: ACPI BIOS Error (bug): Failure creating named object [\_SB.PCI0.GPP8._DSM], AE_ALREADY_EXISTS (20190703/dswload2-327)
juni 24 00:52:21 tux kernel: ACPI Error: AE_ALREADY_EXISTS, During name lookup/catalog (20190703/psobject-221)

Any ideas?
The BIOS update seems to be AMD only stuff. _You know what this is?_
After install it said it was updating the "Led firmware" and nothing more before booting
Changelog:

PRIME X570-PRO BIOS 2203
Update AMD AM4 AGESA PI 1.0.0.1

Damn, I should have searched among our exchanges before upgrading my BIOS.
I'm now stopped to downgrade at version 2206:
Kernel is logging those interrupt handlers errors.
The only workaround is the legacy IRQ way.

Those firmware guys, probably at ASUS; aren't they testing their work with Linux ! At least, take one minute to run Linux from any live media (which will not mess up with their Windows Gamers OS)

What kind of Q&A is made of their department ?

cyring on 12 Nov 2020

I noticed exact same "irq handler" logged after the very first 2XXX BIOS. I then could downgrade the the previous 1407. I tried that just for curiosity some days back. Says the file isn't a BIOS file. What I have not tried yet is to revert back to a previous 2XXX BIOS (if it allows it), either straight to the first or gradually, then to 1407 again.

Though I don't see any problems, I'm hoping with every kernel/firmware/microcode update it will go away. So far not.

I see what's out now is Versjon 2812 Beta (highlighted in Red so I won't try that), which has the AGESA V2 PI 1.1.0.0 Patch C.

I might try the mentioned gradually going back 2XXX if it allows it, if you haven't already?

Though I don't see any problems with it for actual use, not CoreFreq either, so 🤷‍♂️ for me. Just possibly better CPU performance. Also, when I updated to the current 2802 I noticed that strangely the possibility of doing so by Internet in UEFI was gone, no matter if I enabled the Network Stack myself which it always previously did before a reboot into UEFI again and directly into the EZ Flash. And then mountain of new tuning Settings was the real surprise.

I haven't bothered to search but since this appear as a kernel message even with boot parameter quiet it kinda must be reported to kernel devs months ago. First that caused this was first 2XXX which was 2203 for my Motherboard back in June.

On another topic the latest (pretty recent) firmware upgrade to my MP600 NMVes now gives a very steady 5.5 GB/s. Well above specs.

olejon on 12 Nov 2020

With 2311 (patch B), it was so unstable, that I don't give a chance to 2402

I also have to gradually downgrade, version per version.

Interrupt errors are also affecting FreeBSD, according to its mailing list.

cyring on 12 Nov 2020

Bing! Another version arrived today.
Version 2502 has been tried by a few.
Same patch C

cyring on 13 Nov 2020

👍1

Same IRQ errors in linux though 👎 (running Archlinux default kernel)

Got a couple of others to test but these seems to be common across any configuration (Zen, or -ck or -gc).

inglor on 13 Nov 2020

👍1

Bing! Another version arrived today.
Version 2502 has been tried by a few.
Same patch C

Yeah I wrote that about mine but it's beta, no? Any fixes?

olejon on 13 Nov 2020

@inglor @olejon: Posting the issue at ASUS forum

cyring on 13 Nov 2020

Corefreq: Daemon Segmentation Fault on X570 + Ryzen 3

All 154 comments

Linux

Windows

Linux

Windows

Ryzen Master

CoreFreq

Any big difference? Seems to be the top ones under heatsink and the side ones on both next to the side heatsink (and some next to RAM?)

ASUS PRIME X570-PRO

ASUS ROG Crosshair VIII Hero

Idle Temperature - This is just 4 degrees above Room Temperature!

Conic Temperature - Approx 2 minutes - It never goes above 70

More benchmarks and comparisons here:

Almost Idle

During Blender Benchmark

Both Red Rectangles say the same

Screenshots from 2 Manuals for ASRock MBs. Not much info:

Screenshots from 2 Manuals for ASRock MBs. Not much info:

With both set to Auto

Set to Enabled and 100MHz

This is UEFI setting alone causes that big difference in RM's Max Values, not that they are reached at all

Linux

Windows

Power tuning

BIOS settings: all [AUTO] beside a few

Day to day working state

Temperature stays at 50°C

Temperature eventually reaches 80°C

Idle less than 1V - 33°C with Ambient 27°C

Manual OC

BIOS: we're disabling CPB, PBO, XFR

Doing the OC manually

Idle state

Experimental mode

Windows - Corsair MP600 - Basically the first proper PCIe4

Introduction

Windows - Run 1

Windows - Run 2

Windows - Run 3

Windows - Run 4

Windows - Run 5

Linux Clone of same Settings as in Run 3

Linux Clone of same Settings as in Run 5

UEFI AMD OC Built-In PBO Settings

UEFI AMD OC Built-In RAM Settings

UEFI Manual OC Settings

CONCLUSION

Impossible to replicate Run 3 on Windows

The GPU vertical mounting _trend_

* Both those show the same, and the same as in CoreFreq, although it has 1 decimal and CoreFreq *rounds that value Up

* My PC must be on 24/7 now without time for any system "locks" as you asked me to test... No freezing thanks. It has never happened in fact

Related issues

With both set to `Auto`

Set to `Enabled` and `100MHz`

BIOS settings: all `[AUTO]` beside a few

Idle less than `1V` - 33°C with Ambient 27°C

UEFI `AMD OC` Built-In PBO Settings

UEFI `AMD OC` Built-In RAM Settings

UEFI `Manual OC` Settings

* Both those show the same, and the same as in CoreFreq, although it has 1 decimal and CoreFreq rounds that value Up*