I have an FPGA card that sends UDP packets over a 100G Ethernet link. On the receiving side is a ConnectX-6 VPI card an NVIDIA GPU.
Is it possible to use GPUDirect RDMA to capture the UDP packets directly to GPU memory? How do I get started? All of the examples/documentation I’ve seen relate to RoCE or InfiniBand. Can I use GPUDirect RDMA to receive plain UDP packets?
GPUDirect RDMA is an API between IB CORE and peer memory clients.
It provides access to the NIC’s read/write peer memory data buffers, as a result it allows RMDA-based applications to use the peer device without the need to copy data to host memory.
RDMA is based on OS memory management, as it requires pin memory of the OS.
If you want use NIC direct access GPU on the same node to offload CPU on MPI, it requires openmpi and cuda.