NVIDIA/nvshmem

NVIDIA NVSHMEM is a parallel programming interface for NVIDIA GPUs based on OpenSHMEM. NVSHMEM can significantly reduce multi-process communication and coordination overheads by allowing programmers to perform one-sided communication from within CUDA kernels and on CUDA streams.

C++

596

1,238 commits

updated Oct 3, 2026

docs.nvidia.com/nvshmem/api/index.html

See the code

README

NVSHMEM

GPU-initiated communication for scalable NVIDIA GPU clusters.

Introduction

NVSHMEM is an OpenSHMEM-based parallel programming interface that provides a partitioned global address space across NVIDIA GPUs. Applications use symmetric memory for one-sided transfers, atomics, signaling, synchronization, and collective operations. NVSHMEM provides host, CUDA kernel, and CUDA stream interfaces, allowing CPU or GPU threads to initiate communication.

Quick Start

Prerequisites

A Linux system, NVIDIA driver, and CUDA Toolkit supported by the current NVSHMEM release.
Two NVIDIA data center GPUs.
A compatible process launcher, such as Open MPI, Hydra, or Slurm.

Install NVSHMEM

Download a package from the NVSHMEM product page and follow the installation guide. Set NVSHMEM_PREFIX to the installation directory:

export NVSHMEM_PREFIX=/path/to/nvshmem
export LD_LIBRARY_PATH="$NVSHMEM_PREFIX/lib${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}"

Verify the Installation

Launch two NVSHMEM processing elements (PEs), one per GPU, using a process launcher available on your system:

# Open MPI
NVSHMEM_BOOTSTRAP=MPI \
mpirun -np 2 \
  "$NVSHMEM_PREFIX/bin/perftest/device/pt-to-pt/shmem_put_bw" \
  --min_size 8 --max_size 1M --use_smem 0

# Hydra (PMI-1)
NVSHMEM_BOOTSTRAP_PMI=PMI \
mpiexec.hydra -n 2 -ppn 2 \
  "$NVSHMEM_PREFIX/bin/perftest/device/pt-to-pt/shmem_put_bw" \
  --min_size 8 --max_size 1M --use_smem 0

# Slurm (PMI-2)
NVSHMEM_BOOTSTRAP_PMI=PMI-2 \
srun --mpi=pmi2 --nodes=1 --ntasks=2 --gpus-per-task=1 \
  "$NVSHMEM_PREFIX/bin/perftest/device/pt-to-pt/shmem_put_bw" \
  --min_size 8 --max_size 1M --use_smem 0

A successful run reports both GPUs, prints put bandwidth for a range of message sizes, and exits with status zero. If initialization fails, rerun with NVSHMEM_DEBUG=INFO and consult the launching guide. For further help, use the channels in Support and Community.

Build from Source

See the installation guide for source prerequisites and build options. A standard CMake build is:

git clone https://github.com/NVIDIA/nvshmem.git
cd nvshmem

export NVSHMEM_PREFIX="$PWD/install"

cmake -S . -B build \
  -DNVSHMEM_PREFIX="$NVSHMEM_PREFIX" \
  -DCMAKE_INSTALL_PREFIX="$NVSHMEM_PREFIX"
cmake --build build --parallel
cmake --install build

The development branch may be newer than the latest published release and its documentation. For a release-aligned build, check out the corresponding branch or tag before running CMake.

Use NVSHMEM

Installed packages provide imported CMake targets for the host and device libraries:

find_package(NVSHMEM REQUIRED CONFIG)

target_link_libraries(my_target PRIVATE
  nvshmem::nvshmem_host
  nvshmem::nvshmem_device
)

Applications that use only host-side or stream-ordered APIs can link only the host target. See the Using NVSHMEM guide for the programming model, application lifecycle, compilation, and launch requirements. The put-block example provides a complete CUDA program.

Python users should continue with the NVSHMEM4Py quick start.