ARM GPU AMI for AWSBatch/ECS for G5G instance type

This topic has also been touched on here: https://github.com/orgs/aws-samples/projects/55/views/1?pane=issue&itemId=44122312

We would love to hear from you, thank you for using AWS Batch!

You're correct that AWS Batch currently has limited support for ARM-based GPU instances like G5g. While AWS Batch supports GPU jobs using NVIDIA GPUs from various instance families, the G5g instance family (which uses ARM64 architecture) is only supported for GPU jobs when using custom AMIs that you provide to AWS Batch, along with EC2 user data to access the GPUs through customized code and configurations.

At this time, there isn't a pre-built AWS-provided ECS GPU-optimized AMI specifically for ARM-based GPU instances like G5g. To use G5g instances for GPU workloads on AWS Batch with ECS, you would indeed need to build and maintain your own custom AMI that includes the necessary ARM-compatible GPU drivers and configurations.

I cannot provide information about future AWS product roadmaps or whether AWS plans to release an official ARM GPU-optimized AMI for ECS. For the most current information about upcoming features or to express interest in this capability, I'd recommend reaching out through AWS Support channels or your AWS account team, as they can better communicate customer needs to the relevant product teams.

In the meantime, if you want to use G5g instances, you'll need to create a custom AMI with the appropriate ARM-compatible GPU drivers and runtime configurations to enable GPU access for your containerized workloads.
Sources
Run GPU jobs - AWS Batch

We have a AWS DLAMI (Deep Learning AMI) for G5g (ARM64 based Graviton + GPU) instance which should serve as a base image here. Please refer to the following page for where to get the G5g DLAMI from and how to use it. https://docs.aws.amazon.com/dlami/latest/devguide/tutorial-arm64-pytorch.html

Comment added to https://github.com/aws-samples/aws-batch-with-ffmpeg/issues/14

We have a AWS DLAMI (Deep Learning AMI) for G5g (ARM64 based Graviton + GPU) instance which should serve as a base image here. Please refer to the following page for where to get the G5g DLAMI from and how to use it. https://docs.aws.amazon.com/dlami/latest/devguide/tutorial-arm64-pytorch.html