The status of Tensorflow on Arm-based HPC
summary
TensorFlow compiles and runs on AArch64 but only with precise dependency versions, and migrating from TensorFlow 1.15 → 2.2 exposes compatibility and benchmarking gaps; Bazel mismatches (currently using Bazel 0.89 vs 2.2.0) block many official builds. Work targets real server silicon — Ampere Altra / ThunderX3 / A64FX — with SVE/SVE2 optimizations available in GCC/LLVM, OpenCV/OpenCL/OpenMP tuning pending, Python3 CI issues ongoing, a scalable ARM64 colocation cluster is being provisioned for tests, and no official TensorFlow 2.x Docker image for ARM64 exists yet.
AI-written summary · about
chapters (10)
- 0:00Intro
- 0:24Welcome!
- 2:20What is TensorFlow?
- 6:42Versions
- 8:54Build Status
- 13:15Al on Servers Direction
- 17:24Accelerating existing training methods
- 19:48Test Environment
- 20:48Accelerate Deployment of ARM into Data Centers
- 21:45Technology Progress
related talks
43mTensorflow on open source GPUs
36mWhy and How to Choose Arm64 for your Cloud Applications
52mUbuntu on ARM
29mThe Progress for Cloud Computing on Arm architecture
44mWhat to know before adopting Arm: Lessons learned at Datadog
23mMulti-Arch Support in OpenStack
30mApache projects, let's embrace ARM!
31mOptimizing Cost and Performance with arm64