Question
We are testing NCCL GIN on machines equipped with ConnectX-6 (CX6) NICs. The GIN initialization completes without errors (doca_gpu_create, QP creation, and QP state transitions all succeed), but at runtime, GPU-initiated RDMA operations (both put_signal and signal) silently fail — the WQE appears to be posted, yet the remote signal is never incremented, causing waitSignal to hang indefinitely. No errors are reported even with DOCA_GPUNETIO_LOG=6.
The same tests pass on ConnectX-7 (CX7) NICs. We have also verified that CX6 hardware does support GPU-initiated RDMA (including atomics) through a separate DevX-based implementation.Could you confirm whether CX6 is a supported NIC for GIN / GDAKI, or if CX7+ is a hard requirement?
Question
We are testing NCCL GIN on machines equipped with ConnectX-6 (CX6) NICs. The GIN initialization completes without errors (doca_gpu_create, QP creation, and QP state transitions all succeed), but at runtime, GPU-initiated RDMA operations (both put_signal and signal) silently fail — the WQE appears to be posted, yet the remote signal is never incremented, causing waitSignal to hang indefinitely. No errors are reported even with DOCA_GPUNETIO_LOG=6.
The same tests pass on ConnectX-7 (CX7) NICs. We have also verified that CX6 hardware does support GPU-initiated RDMA (including atomics) through a separate DevX-based implementation.Could you confirm whether CX6 is a supported NIC for GIN / GDAKI, or if CX7+ is a hard requirement?