OpenGL on AWS g5g.metal (arm64) instance with docker

I’m trying to run an OpenGL workload on a g5g.metal instance. These are arm64 instances with Nvidia T4G graphic cards.

I have deployed a g5g.metal instance with using the NVIDIA GPU-Optimized AMI (ARM64). This comes with docker preinstalled.

Because I have built a docker instance that is using a newer version of CUDA as described here, I had to upgrade the nvidia driver on the machine to nvidia-driver-535-server.

Then I have tried running the docker image that I have built like this:

$ docker run --rm -it --entrypoint /bin/bash --gpus all -e NVIDIA_VISIBLE_DEVICES=all -e NVIDIA_DRIVER_CAPABILITIES=all mycudagl:12.2.2-runtime-ubuntu22.04

Then, inside the running container:

# ls /lib/aarch64-linux-gnu/*GL*
/lib/aarch64-linux-gnu/libEGL_nvidia.so.0           /lib/aarch64-linux-gnu/libGLESv1_CM_nvidia.so.1           /lib/aarch64-linux-gnu/libGLESv2_nvidia.so.2           /lib/aarch64-linux-gnu/libGLX_indirect.so.0  /lib/aarch64-linux-gnu/libGLX_nvidia.so.535.161.07
/lib/aarch64-linux-gnu/libEGL_nvidia.so.535.161.07  /lib/aarch64-linux-gnu/libGLESv1_CM_nvidia.so.535.161.07  /lib/aarch64-linux-gnu/libGLESv2_nvidia.so.535.161.07  /lib/aarch64-linux-gnu/libGLX_nvidia.so.0

Note that it did not show libEGL.so.1 nor libGLESv2.so.2 which my OpenGL application relies on.

I have ran the same container on my machine (but the x86_64 version on GeForce GTX 1050 Ti, Driver Version: 545.29.06, CUDA Version: 12.3) and the above mentioned libraries are there.

Is arm64 not fully supported with nvidia docker or am I doing something wrong?