NVIDIA HGX B200/B300 GPU Baseboard
In the NVIDIA HGX B300/B200/B100 baseboard PCIe topology, NVSwitches are not recognized as PCIe devices on the host system. To manage NVLinks, the NVSwitches are connected to a CX7 Bridge device. The host stack and control plane access are routed through this CX7 bridge device and use the corresponding in-box OFED or the MOFED driver.
The CX7 bridge device is integrated into the GPU baseboard, which includes two physical ports. Each port exposes one physical function (FC PF) and one Limited physical function (LPF) to the host system, which totals four PFs. The PFs are categorized into the following PFs:
-
Limited PFs (LPF) are designated for specific tasks in the system.
They are used by the FM and the NVLSM to configure and set up NVSwitches, GPU, and NVLink routing information. LPFs are also used by telemetry agents, such as NVIBDM and DCGM, to monitor and collect data. Resetting this PF with FLR also resets the corresponding NVSwitch device.
-
Full Capabilities PF (FC PF) provides device administration level functionalities, such as issuing NVSwitch device resets and enabling or disabling links between NVSwitches.
FC PFs are valuable for partial pass-through virtualization scenarios, where subsets of GPUs and NVSwitches are allocated to Tenant VMs. However, this PF type does not support NVLink control plane entities, such as FM and NVLSM, and communication with telemetry agents.
From a hardware perspective, the CX7 Bridge device for NVLink management and traditional CX7 NICs share identical hardware. The PCIe Vital Product Data (VPD) information, which is programmed during production, differentiates the CX7 device for NVLink Management:
-
On Linux-based systems, VPD information can be accessed using standard tools such as
vpddecodeorlspcicommands. -
On Windows-based host systems, to query VPD information, run the
mstregcommand.
To differentiate between LPFs and FC PFs, the LPF VPD information includes a vendor-specific field called SMDL, with a non-zero value defined as SW_MNG. For bare-metal, full pass-through, and shared NVSwitch deployments, the prelaunch script in the FM service unit file will run and query the available CX7 devices for this VPD information. The file populates the required FM and NVLSM configuration values so that these communication entities can access the relevant devices.
However, for partial pass-through deployments, additional steps are required at the hypervisor level to identify the pair of PFs that belong to a CX7 bridge device port. Refer to Virtualization Models for more information.
The NVIDIA HGX B300/B200/B100 baseboard PCIe topology will display eight GPUs and four CX7 devices on the PCIe tree as PCIe devices on the host system.
Here is an example:
lspci | grep -i -E 'nvidia|mella'
05:00.0 Infiniband controller: Mellanox Technologies MT2910 Family [ConnectX-7]
05:00.1 Infiniband controller: Mellanox Technologies MT2910 Family [ConnectX-7]
05:00.2 Infiniband controller: Mellanox Technologies MT2910 Family [ConnectX-7]
05:00.3 Infiniband controller: Mellanox Technologies MT2910 Family [ConnectX-7]
1b:00.0 3D controller: NVIDIA Corporation Device 29bc (rev a1)
43:00.0 3D controller: NVIDIA Corporation Device 29bc (rev a1)
52:00.0 3D controller: NVIDIA Corporation Device 29bc (rev a1)
61:00.0 3D controller: NVIDIA Corporation Device 29bc (rev a1)
9d:00.0 3D controller: NVIDIA Corporation Device 29bc (rev a1)
c3:00.0 3D controller: NVIDIA Corporation Device 29bc (rev a1)
d1:00.0 3D controller: NVIDIA Corporation Device 29bc (rev a1)
df:00.0 3D controller: NVIDIA Corporation Device 29bc (rev a1)
The GPU and NVSwitch PCIe Device ID, Product ID information in the lspci command output above is only provided as a sample. Actual values might vary depending on the specific product and GPU variations in the system configuration.

浙公网安备 33010602011771号