Science Goals

AI at scale is communication-bound, and the software beneath it was not designed for the job.

• MPI carries decades of portable, specified semantics, but predates modern AI collectives and GPU-resident data.
• NCCL and its ports are vendor toolkits, not specifications. One vendor sets the pace; every other ∗CCL reimplements its API.
• AI has no vendor-neutral, resilient, GPU-native communication substrate.

Aim

Give AI workloads the portability MPI already gives simulation, without surrendering the performance frameworks buy today through lock-in.

Objectives

• improve MPI for AI
• improve Open MPI
• contribute upstream
• devise HPC+AI APIs
• unify MPI with NCCL and comparable platforms.