I have a few technical questions regarding the RDMA capabilities of the RTX 6000 Ada Generation series:
RDMA Support & MIG: Does the RTX 6000 series officially support GPUDirect RDMA? If so, is it mandatory to enable MIG (Multi-Instance GPU) to utilize RDMA features?
CPU Bottleneck Issues: I have observed some demonstrations where data appears to pass through the CPU even when RDMA is enabled, causing performance bottlenecks. If the RTX 6000 supports RDMA, does it suffer from this same architectural limitation?
P2P Communication: When using GPUDirect RDMA, can we achieve pure P2P (Peer-to-Peer) communication that completely bypasses the CPU for data transfers between GPUs or network interfaces?