GCP now auto-loads the ConnectX NCCL plugin
- Source
- x.com
- Date
Amazing work by NVIDIA & Google on implementing auto-activation out of the box for Google's NCCL Plugin for ConnectX-7/8 NCCL. This was implemented due to feedback from SemiAnalysis a couple of quarters ago to massively improve the out-of-the-box experience on the GCP platform! Previously, users would need to mess around with the correct library load paths and env vars to get it set up correctly for optimized performance on NVIDIA ConnectX NICs on GCP NVIDIA GPU machines, but now it is fully automated!
Context
NVIDIA's PyTorch container can activate Google Cloud's GPU-communication plugins automatically on supported machines. Its 26.07 release notes specify A3 Ultra, A4, A4X and A4X Max hosts, where the container detects the platform and enables the relevant networking and tuning components.
This puts configuration needed for communication between GPU machines into the container's startup behavior. SemiAnalysis describes the change as removing manual library-path and environment-variable setup. The supported change is easier activation in that container and hardware combination; neither source supplies a measured speedup from automatic activation itself.
Checking sign-in…
Loading comments…