Hi NVIDIA team,
I am looking for public guidance on how to enable or use end-host multipath RDMA on ConnectX-8, including MRC and the DOCA-PRDMA interface mentioned in recent public materials.
I saw the recent MRC paper, “Resilient AI Supercomputer Networking using MRC and SRv6”, which mentions that MRC has been implemented on NVIDIA ConnectX-8 RDMA NICs:
I also saw the ConnectX-8 Hot Chips 2025 slides, which mention integrated load balancing/congestion control, and the software APIs “DOCA-PRDMA”:
However, I could not find public documentation for either of the following:
- How to enable or use MRC on ConnectX-8
- Any public DOCA-PRDMA documentation, headers, samples, or programming guide
Could you clarify the current support status?
Specific questions:
- Is MRC available to external customers on ConnectX-8 today?
- If yes, what are the required firmware, MLNX_OFED / inbox driver, rdma-core, and DOCA versions?
- Is DOCA-PRDMA publicly available in any DOCA release, or is it currently limited to early-access / NDA programs?
- If DOCA-PRDMA is not public yet, is there a recommended public API for experimenting with programmable RDMA transport behavior or load-balancing policy on CX8?
Thanks!