NVIDIA HPC-X
Type: Library Tags: NVIDIA, HPC-X, MPI, SHMEM, PGAS, UCX, UCC, HCOLL, SHARP, InfiniBand, Spectrum-X, NCCL Related: NVIDIA-HPC-SDK, NVIDIA-HPC-Compilers, NVIDIA-Quantum-InfiniBand, NVIDIA-Spectrum-X, NVIDIA-ConnectX-InfiniBand, NVIDIA-DOCA, NVIDIA-DOCA-OFED, NVIDIA-MLNX-OFED, NCCL, NVSHMEM, GPUDirect-RDMA Sources: https://docs.nvidia.com/networking/display/hpcxv226; https://developer.nvidia.com/networking/hpc-x Last Updated: 2026-04-29
Summary
NVIDIA HPC-X is NVIDIA’s communications toolkit for high-performance MPI, SHMEM, and PGAS applications on NVIDIA networking. The current docs line is HPC-X Rev 2.26, tested as part of Spectrum-X reference architecture 2.1, and includes MPI, OpenSHMEM, UCX, UCC, HCOLL, ClusterKit, NCCL-RDMA-SHARP plugins, multi-GPU support, and Spectrum-X NCCL plugin guidance. It is the main NVIDIA networking software bridge between traditional HPC communication models and GPU-heavy AI/HPC clusters.
Detail
HPC-X packages and optimizes the communications libraries that many HPC and AI applications rely on. NVIDIA describes it as a comprehensive software package including MPI and SHMEM communications libraries plus acceleration packages that improve performance and scalability.
The toolkit includes HPC-X MPI, an optimized Open MPI implementation; HPC-X OpenSHMEM for one-sided communication; Unified Communication X (UCX) for low-overhead point-to-point communication; Unified Collective Communication (UCC) for collectives; HCOLL for collective acceleration; ClusterKit for cluster health and performance assessment; and NCCL-RDMA-SHARP plugins that integrate NVIDIA collective communication with RDMA and SHARP.
HPC-X is adjacent to, but distinct from, NCCL and NVSHMEM. NCCL is the deep-learning collective library, NVSHMEM is a GPU-cluster PGAS programming model, and HPC-X provides the broader MPI/SHMEM/UCX/UCC toolkit used by HPC applications and cluster validation. All three depend on the same NVIDIA fabric strengths: NVIDIA-Quantum-InfiniBand, NVIDIA-ConnectX-InfiniBand, GPUDirect-RDMA, SHARP, and increasingly NVIDIA-Spectrum-X for Ethernet AI networking. On hosts, the relevant NVIDIA driver substrate is now NVIDIA-DOCA-OFED, with NVIDIA-MLNX-OFED remaining important for legacy LTS deployments.
Connections
- NVIDIA-Quantum-InfiniBand - HPC-X is optimized for Quantum InfiniBand networking and SHARP in-network computing.
- NVIDIA-HPC-SDK - current HPC SDK docs list HPC-X MPI as the SDK’s MPI communication library surface.
- NVIDIA-HPC-Compilers - compilers and MPI wrappers often meet in HPC-X-style HPC build environments.
- NVIDIA-Spectrum-X - HPC-X 2.26 is tested with Spectrum-X reference architecture 2.1 and includes Spectrum-X NCCL plugin documentation.
- NVIDIA-ConnectX-InfiniBand - ConnectX adapters provide the RDMA transport hardware used by HPC-X libraries.
- NVIDIA-DOCA - DOCA and DOCA-OFED provide the host/networking software substrate for newer NVIDIA network deployments.
- NVIDIA-DOCA-OFED - current Linux host-driver stack for NVIDIA RDMA/RoCE networking.
- NVIDIA-MLNX-OFED - legacy standalone Linux OFED stack used by older HPC-X deployments and runbooks.
- NCCL - HPC-X includes NCCL-RDMA-SHARP plugin coverage and complements NCCL collective communication.
- NVSHMEM - HPC-X includes SHMEM/PGAS communications that sit near NVSHMEM’s GPU-focused PGAS model.
- GPUDirect-RDMA - HPC-X uses GPU-aware RDMA paths for CUDA-enabled communication workloads.
Source Excerpts
- “HPC-X v2.26 is tested as part of Spectrum-X reference architecture release version 2.1.”
- “HPC-X also includes various acceleration packages.”