Hardware and software requirements

cuFFTMp relies on NVSHMEM for GPU-initiated communications. It inherits many of its requirement.

Hardware requirements

  • x86_64 or ppc64le CPU architecture

  • NVIDIA data center GPUs, of the SM70 (Volta) or SM80 (Ampere) architecture

  • All GPUs should be

    • peer-to-peer connected

    • or connected via GPUDirect RDMA over Infiniband using a Mellanox adapter (CX-4 or later)


Note that being merely connected via PCIe does not guarantee peer-to-peer access between GPUs. Also note that NVLink is not required for peer-to-peer access, but it is the preferred transport.

Software requirements

cuFFTMp is shipped as part of the HPC SDK 22.3 and requires

  • Linux 64 bits

  • On x86_64, CUDA 11.2 or later. On ppc64le, CUDA 11.4 or later.