This guide is a platform-neutral reference for triaging standalone NVIDIA® ConnectX® network interface cards (NICs). It covers Ethernet, RDMA over Converged Ethernet (RoCE), and InfiniBand.
This guide combines validated NVIDIA guidance on SOS and MFT collection, PCIe escalation, ConnectX diagnostics, physical-link health, and InfiniBand support with up-to-date performance-troubleshooting references.
BlueField and Platform-specific Applicability
For BlueField operating in NIC mode, use this guide for ConnectX networking triage. For issues involving BlueField DPU mode, Arm cores, or DPU services, refer to the NVIDIA BlueField Platform Software Troubleshooting Guide.
For platform-specific procedures on NVIDIA DGX GB200/GB300 NVL72 systems, refer to the NVIDIA DGX GB Rack Scale Systems User Guide. For other systems, follow the system vendor’s documentation and support procedures.
Working Principles
-
Start with the observed symptom and the exact error.
-
Collect evidence before you reboot or reset the device.
-
For intermittent issues, keep evidence from both a good state and a bad state.
-
Change one component or setting at a time, and record the result of each change.
-
Confirm root cause with a controlled comparison. Recovery after a reboot, reseat, or swap shows only that the failure is intermittent or that a component is suspect.
-
Keep customer steps separate from internal-only diagnostics. Never ask customers to run internal-only tools.
-
Follow the decision path for the symptom instead of running every available command.
Guide Structure
Document Revision History
Refer to Document Revision History.
Last updated: