Is this page helpful?

Known Limitations#

  • Live migration is not supported in GPU passthrough configurations since the host CPUs, hugepage-backed NUMA memory, 8 x GPUs and the NVSwitch fabric are tied to the physical server.

  • Suspend, hibernate and managed-save operations are not supported with GPU passthrough. Shut down the guest before returning devices to the host.

  • The host and guest cannot manage the GPU fabric at the same time. Keep the host GPU Operator components disabled (nvidia.com/gpu.deploy.operands=false) while the guest owns the GPUs and bridge functions, and do not start the guest while host workloads use the GPUs.

  • Host BDFs are system-specific. Guest BDFs are generated from the virtual PCIe hierarchy. Do not reuse this device map on another B300 system without confirming its topology.

  • The guest reserves 120 of the host’s 128 physical cores and 1.5 TiB of its 2 TiB memory. Account for the reserved host resources when comparing performance.

  • Virtual root-port information does not show the physical PCIe endpoint link. Check the host endpoint, firmware, Fabric Manager, and workload behavior when troubleshooting PCIe performance.

  • For supported NVSwitch virtualization models and B300 bridge-device assignment, please refer to the NVIDIA Fabric Manager User Guide.