Pytorch_geometric: Hardware for PyTorch Geometric

Created on 15 Aug 2019  路  1Comment  路  Source: rusty1s/pytorch_geometric

This paper got me wondering if there is any specific type of GPU or similar that works particularly well with this software / with GNNs in general. What do you recommend / what has your experience been? How does the conclusion of this blog post apply to PyTorch Geometric?

picture testzorg

Most helpful comment

Thank you for the link to the paper. I think that the conclusion of the blog post does also hold for sparse architectures like they are implemented in PyTorch Geometric. At work I have access to a 1080Ti ans 2080Ti and both work great, but comparably equal. There are also a lot of students of mine which run PyG on lower hardware and they never complained. I do not have access to any other setups, so it is quite hard to answer your question, e.g., I never played around with TPUs, but according to this paper GNNs can also be heavily accelerated by the use of TPUs and a clever memory layout.

picture rusty1s on 15 Aug 2019
👍3

>All comments

Thank you for the link to the paper. I think that the conclusion of the blog post does also hold for sparse architectures like they are implemented in PyTorch Geometric. At work I have access to a 1080Ti ans 2080Ti and both work great, but comparably equal. There are also a lot of students of mine which run PyG on lower hardware and they never complained. I do not have access to any other setups, so it is quite hard to answer your question, e.g., I never played around with TPUs, but according to this paper GNNs can also be heavily accelerated by the use of TPUs and a clever memory layout.

rusty1s picture rusty1s on 15 Aug 2019
👍3
Was this page helpful?
0 / 5 - 0 ratings

Related issues

FerranAlet picture FerranAlet  路  4Comments
yanzhangnlp picture yanzhangnlp  路  3Comments
zhangfuyang picture zhangfuyang  路  4Comments
Zhangzhk0819 picture Zhangzhk0819  路  3Comments
WMF1997 picture WMF1997  路  4Comments