Hardware and software requirements

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

Hardware requirements

  • x86_64 or ppc64le CPU architecture

  • NVIDIA data center GPUs, of the Volta, Ampere or Hopper architecture

  • All GPUs should be

    • peer-to-peer connected

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

Warning

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.

Warning

On aarch64, peer-to-peer is supported on Grace-Hopper. On non Grace-Hopper platforms, peer-to-peer is supported only for NVLink-connected GPUs or PCIe-connected GPUs using a PCIe switch.

Software requirements

cuFFTMp is shipped as part of the HPC SDK and requires

  • Linux 64 bits

  • A CPU with x86_64 or aarch64 architecture

  • CUDA 11.0 or later