Copilot-cli: GPU Support

[hosymia0]

Does Copilot currently support GPU based task definitions? Similar to https://docs.docker.com/cloud/ecs-integration/?

[bvtujo]

Copilot uses ECS Fargate under the hood, which does not yet support GPUs. You can track ECS' progress on this issue over at the Containers Roadmap. We have plans to enable this once it's available depending on customer use cases and feedback!

[efekarakus]

+1 to Austin's reply :) @hosymia0 if you don't mind can you tell us the architecture of services that would use GPUs?

Do you expect this service to consume events from a SQS queue? or will this service be triggered by a Load Balancer? or are they just one-off tasks that need GPUs?

[neosavvy]

+1 to Austin's reply :) @hosymia0 if you don't mind can you tell us the architecture of services that would use GPUs?

Do you expect this service to consume events from a SQS queue? or will this service be triggered by a Load Balancer? or are they just one-off tasks that need GPUs?

I also would be very interested in evaluating this as an option.

I have SciKit Learn code that would run in milliseconds on a GPU enabled machine (via CUDA).

It needs to execute on incoming web requests and for a new service they will be intermittent. Then on a schedule (nightly or weekly) we will reprocess all of those models and notify users.

[efekarakus]

Ok awesome thanks for the input @neosavvy, it's super helpful !
Sounds like a gpu field would be useful for both existing service types: "Load Balanced Web Service" and "Backend Service", as well as for the "Scheduled Job" option to train models periodically.