VF Network Function Limitations in SR-IOV Legacy Mode & in Switchdev Mode
|
Dual Port Device |
Single Port Device |
|---|---|
|
127 VF per PF (254 functions) |
127 |
VF+SF Network Function Limitations in Switchdev Mode
|
Dual Port Device |
Single Port Device |
|---|---|
|
|
|
Internal ref. |
Issue |
|---|---|
|
5236199 |
Description: Due to a limitation in module telemetry retrieval, pre-FEC BER and related telemetry values remain at zero and do not transition to live telemetry status. Other telemetry defined by the CMIS VDM specification is expected to function normally. |
|
Workaround: N/A |
|
|
Keywords: Telemetry retrieval, pre-FEC BER |
|
|
Detected in version: 40.51.1002 |
|
|
5329699 / 5388950 |
Description: Setting adp_retx_profile_id to 0 through ROCE_ACCL after a user profile has been configured is silently ignored. |
|
Workaround: N/A |
|
|
Keywords: ROCE_ACCL, adp_retx_profile_id |
|
|
Detected in version: 40.51.1002 |
|
|
301219 / 5224356 |
Description: On GB200 setups with DDP enabled, the HPC-X application hangs and prints the warning "got active message id 0, but no handler installed". |
|
Workaround: Set the environment variable UCX_IB_ODP_MEM_TYPES=host. |
|
|
Keywords: GB200, DDP, HPC-X, UCX, hang |
|
|
Detected in version: 40.51.1002 |
|
|
5043315 |
Description: When running bidirectional ib_write_bw traffic on both interfaces simultaneously, low performance may be observed due to high PCIe latency (~4 µs) in certain system configurations. |
|
Workaround: Decrease GABL to 192 KB (default is 282 KB for this setup) using mlxconfig to improve bidirectional bandwidth to approximately 1,450 Gb/s. |
|
|
Keywords: ib_write_bw, PCIe latency, GABL, mlxconfig, bidirectional bandwidth |
|
|
Detected in version: 40.51.1002 |
|
|
4893565 / 4532684 |
Description: Occasional high spikes in operation latency may be observed. |
|
Workaround: N/A |
|
|
Keywords: latency spike, operation latency |
|
|
Detected in version: 40.51.1002 |
|
|
4759910 / 5011023 |
Description: During an ISSU or Live Firmware Patch over InfiniBand, the NIC may be temporarily unreachable by the OpenSM subnet manager for the duration of the ISSU process, which can result in "Failed to get sminfo" errors. |
|
Workaround: N/A |
|
|
Keywords: ISSU, LFWP, InfiniBand, OpenSM, sminfo |
|
|
Detected in version: 40.51.1002 |
|
|
4705948 / 4684233 |
Description: A bandwidth degradation of approximately 25% may be observed when performing DC RDMA Read operations between ConnectX-7 (NDR) and ConnectX-8 using a low number of QPs (1–16). The degradation diminishes as QP count increases. |
|
Workaround: N/A |
|
|
Keywords: DC RDMA Read, bandwidth degradation, NDR, QP |
|
|
Detected in version: 40.51.1002 |
|
|
4705241 / 4684233 |
Description: When Quantum-2 is part of an XDR topology, serving as a leaf switch connected to NDR-based hosts, a bandwidth degradation of approximately 3–7 Gb/s is expected. |
|
Workaround: N/A |
|
|
Keywords: Quantum-2, XDR, NDR, leaf switch, bandwidth degradation |
|
|
Detected in version: 40.51.1002 |
|
|
4628696 |
Description: When HASH LAG single QP is enabled, ib_read_bw for a single QP over hash LAG can reach up to 337 Gbps, while ib_write_bw for a single QP can achieve up to 390 Gbps. |
|
Workaround: N/A |
|
|
Keywords: HASH LAG, single QP, ib_read_bw, 337 Gbps |
|
|
Detected in version: 40.51.1002 |
|
|
4436843 |
Description: When the link recovery feature is enabled with a link-down timeout of 0, port recovery counters are unexpectedly reset. |
|
Workaround: N/A |
|
|
Keywords: link recovery, PPRM, recovery counter, link-down timeout |
|
|
Detected in version: 40.51.1002 |
|
Internal Ref. |
Issue |
|---|---|
|
4970263 / 4989004 |
Description: A high-temperature assert may be reported only the first time the chip reaches the thermal threshold. |
|
Workaround: N/A |
|
|
Keywords: Temperature |
|
|
Detected in version: 40.49.1014 |
|
|
5085252 / 5090080 |
Description: Following a hot reset, a false Surprise Link Down (SLD) alarm may be reported intermittently on the Downstream Port (DSP) through the Advanced Error Reporting (AER) register. This is a false alarm and has no functional impact on link stability or data integrity. |
|
Workaround: N/A |
|
|
Keywords: Surprise Link Down (SLD) |
|
|
Detected in version: 40.49.1014 |
|
|
4901395 / 4919262 |
Description: When PXE is enabled for both Legacy (FlexBoot) and UEFI drivers, the UEFI driver supports only a single TFTP session after the DHCP handshake. |
|
Workaround: For scenarios that require multiple concurrent TFTP sessions, disable PXE filters via the HCA HII menu. |
|
|
Keywords: PXE, Filter enabled, UEFI |
|
|
Detected in version: 40.49.1014 |
|
|
4893639 |
Description: Direct Memory Access (DMA) protection is not supported in the ipxe.efi driver. |
|
Workaround: N/A |
|
|
Keywords: DMA |
|
|
Detected in version: 40.48.1000 |
|
|
4594515 / 4622127 / NVbug 5460310 |
Description: Running perftest with ib_send_bw using a message size of 4KB and 1000 QPs does not reach the expected full line rate. |
|
Workaround: N/A |
|
|
Keywords: perftest |
|
|
Detected in version: 40.48.1000 |
|
|
4604969 |
Description: Probe packets might be dropped at the transmission stage when multiple congestion control flows are active. |
|
Workaround: N/A |
|
|
Keywords: PCC, RTT, probe |
|
|
Detected in version: 40.47.1026 |
|
|
4496642 |
Description: The timestamps (t2, t4) of the received RTT probes are taken from the free-running clock, even when ROCE_CC_RTT_TIMESTAMP_FORMAT is set to 0x02.
|
|
Workaround: N/A |
|
|
Keywords: RTT RTC timestamp |
|
|
Detected in version: 40.47.1026 |
|
|
4705241 |
Description: When Quantum-2 is part of an XDR topology, serving as a leaf switch connected to NDR-based hosts, a bandwidth degradation of approximately 3–7 Gb/s is expected. |
|
Workaround: N/A |
|
|
Keywords: XDR, NDR, Quantum-2 |
|
|
Detected in version: 40.47.1026 |
|
|
4705948 |
Description: When using DC as the InfiniBand transport type to perform an ib_read RDMA operation between ConnectX-7 (NDR) and ConnectX-8, a bandwidth degradation of approximately 25% may be observed when using a low number of QPs (1–16). The performance degradation diminishes as the number of QPs increases. |
|
Workaround: N/A |
|
|
Keywords: DC, ib_read RDMA, NDR, performance |
|
|
Detected in version: 40.47.1026 |
|
|
4685736 |
Description: Creating a DPA process that allocates a 128 MB data segment and loads a dynamic library may fail with syndrome 0xdc30ac. |
|
Workaround: Limit the DPA application’s data segment size to 64 MB. |
|
|
Keywords: DPA |
|
|
Detected in version: 40.47.1026 |
|
|
4618452 |
Description: When only a partial number of fiber lanes are connected to a module (i.e., not all supported lanes are populated), the module re-insertion process may experience an extended link-up time of approximately one minute. This occurs because the module requires additional time to detect that only a subset of the lanes is connected. |
|
Workaround: N/A |
|
|
Keywords: Cables |
|
|
Detected in version: 40.47.1026 |
|
|
4270913 |
Description: For DC InfiniBand transport, using SRQs requires distributing QPs across four SRQs. Each QP must be assigned to its own SRQ when there are up to four QPs. |
|
Workaround: If more than four QPs are used, they must be evenly distributed across the four available SRQs. |
|
|
Keywords: DC InfiniBand transport, SRQ, QPs |
|
|
Detected in version: 40.47.1026 |
|
|
4665802 |
Description: Due to a re-burst scheduler limitation, the expected “speed of light” value for single QP RDMA Write on ConnectX-8 devices cannot be calculated. Currently, verification is limited to identifying performance degradations relative to previous firmware versions. |
|
Workaround: N/A |
|
|
Keywords: Single QP RDMA Write |
|
|
Detected in version: 40.47.1026 |
|
|
4394475 |
Description: The existing congestion control configuration applies globally, rather than on a per-priority basis. |
|
Workaround: Ensure that the configuration values for all priorities are aligned in either |
|
|
Keywords: Congestion control, ROCE_CC_PRIO |
|
|
Detected in version: 40.45.1020 |
|
|
4201405 |
Description: Upgrading to firmware 40.44.0xxx from any previous Engineering Sample version requires power cycling the driver and not just resetting it (using mlxfwreset). |
|
Workaround: N/A |
|
|
Keywords: Upgrade, power cycle, reset |
|
|
Detected in version: 40.44.0208 |
Last updated: