The Ray community has announced a Special Interest Group focused on hardware accelerators such as GPUs and TPUs. Ray was created at UC Berkeley over 10 years ago, when distributed computing was built on CPU clusters. Ray already supports more than a dozen kinds of accelerators, and the new group will bring together hardware vendors, Ray contributors, and application developers.
The group aims for a shared, vendor-neutral roadmap. Its first roadmap covers 3 core areas: development and testing infrastructure, hardware-aware scheduling, and networking and I/O transport. Ray also needs an accurate picture of each node's hardware structure, so its scheduler can place work on the right hardware. Direct tensor transfer between devices is also a focus, via Ray Direct Transport and transports such as UCX and RDMA.
Core scheduling and placement logic stay in Ray Core, while the SIG defines plug-in interfaces, CI testing, and maturity tiers for accelerators. Hardware vendors and cloud providers will maintain their own device plugins and transports. People can join the Ray Slack, attend the community meetings to discuss proposals and review roadmap priorities, or check open accelerator issues and pull requests on GitHub.
Key facts
- The Special Interest Group for Ray on accelerators is also called SIG Accelerator.
- The initial roadmap covers 3 core areas of work.
- Ray supports more than a dozen kinds of hardware accelerators today.
- Ray Direct Transport lets actors pass tensor references while data moves directly between devices.
- People can join through the Ray Slack, community meetings, and GitHub issues and pull requests.
Why it matters
For people who use Ray, the group aims to make accelerators easier to use, more efficient at scale, and more reliable in production, with clearer hardware information for placing work. A documented process would bring vendor and cloud accelerators into Ray's testing, giving contributors access to reproduce and debug failures. For those building integrations, the roadmap includes a documented extension point and checklist for adding new accelerators.




