OpenSHMEM's new GPU-aware spec unifies heterogeneous HPC communication

The paper, authored by researchers from institutions including Oak Ridge National Laboratory, AMD, Intel, NVIDIA, and several universities, tackles a critical problem in modern HPC: the lack of portable GPU memory semantics in the OpenSHMEM 1.x standard. As supercomputers shift to accelerator-centric architectures where GPUs provide the majority of floating-point throughput and memory bandwidth, existing GPU-enabled OpenSHMEM implementations have diverged in how they handle memory management, capability discovery, and operation semantics. This fragmentation threatens the community unification that OpenSHMEM was originally designed to achieve.

The proposed auxiliary specification is designed as a minimal, backward-compatible extension that can be rapidly implemented. It introduces a GPU-scoped memory space abstraction, along with well-defined semantics for remote memory access (RMA), atomic operations, synchronization, and collectives when using GPU-attached buffers. The initial focus is on a host-initiated interface, but the general semantics also support optional device-initiated operations. By providing a vendor-neutral specification that can work across NVIDIA, AMD, and Intel GPUs, this effort aims to enable write-once, run-anywhere GPU communication code, accelerating development for HPC applications targeting exascale and beyond.