Q: XLA Linux ARM64 CPU build is based on Ubuntu 20.04

[apivovarov]

I noticed that the XLA Linux ARM64 CPU CI build uses GCC-9 for building XLA on ARM64.

Did small investigation and found the following:

• The XLA Linux ARM64 CPU CI build uses the Docker image us-central1-docker.pkg.dev/tensorflow-sigs/tensorflow/ml-build-arm64:latest.
• This image corresponds to the following TensorFlow Dockerfile:
  https://github.com/tensorflow/tensorflow/blob/master/ci/official/containers/ml_build_arm64/Dockerfile
• The ARM64 ml_build image is based on Ubuntu 20.04 (which includes GCC-9).

In contrast, the x86_64 ml_build image is based on Ubuntu 22.04 (which includes GCC-11):
https://github.com/tensorflow/tensorflow/blob/master/ci/official/containers/ml_build/Dockerfile

Given the discrepancy in base images and GCC versions, should we open a PR to update the ARM64 TensorFlow ml_build_arm64 image to use Ubuntu 22.04 for consistency in Ubuntu and GCC versions?

David, what you think? @ddunl

[ddunl]

I think that we still use clang on the ARM build based on https://github.com/openxla/xla/blob/main/tensorflow.bazelrc#L882, I think in CI at least we always override the default toolchain. I'm open to using a newer ubuntu though, cc: @MichaelHudgins who may have some input