NVSHMEM: Parallel programming interface for NVIDIA GPU clusters

Fields

id

sources/src5-nvshmem-developer-page

URL

https://developer.nvidia.com/nvshmem

Publisher

NVIDIA

Kind

whitepaper

Published

2024-01-01

Accessed

2026-10-04

notes

NVSHMEM is based on OpenSHMEM. Creates global address space across multiple GPUs. GPU-initiated operations, CPU-initiated operations, stream-based operations. v3.8 adds: host/device RMA regions, counted-signal APIs, multicast logical endpoints, CUDA 13.4 features for Vera Rubin. Key for multi-GPU inference serving.

Notes

NVSHMEM is based on OpenSHMEM. Creates global address space across multiple GPUs. GPU-initiated operations, CPU-initiated operations, stream-based operations. v3.8 adds: host/device RMA regions, counted-signal APIs, multicast logical endpoints, CUDA 13.4 features for Vera Rubin. Key for multi-GPU inference serving.

Cross-references

This record references no other records.