KubeVirt on GB200: Virtualizing a Rack-Scale Supercomputer#

Fan Zhang, Kevin Klues & Alay Patel, NVIDIA — Cloud Native Theater | KubeVirt Summit

• Why: The clearest statement in the program of how rack-scale hardware breaks existing software assumptions. When the rack behaves as one logical system, the PCIe-shaped abstractions in the VFIO/QEMU/Kubernetes stack stop describing reality.
• What: The practical enablement path for running KubeVirt on GB200 (Grace Blackwell), where CPU and GPU are coupled through a cache-coherent interconnect with unified memory rather than discrete devices on PCIe.
• How: VFIO and kernel requirements, QEMU/libvirt requirements, Topology Manager requirements for device plugins or DRA, and two rack-scale pieces: Compute Domains as the unit of allocation for multi-node GPU fabrics, with IMEX domain bring-up/teardown orchestrated by a daemon integrated into the KubeVirt lifecycle; and guest topology pass-through mirroring host CPU/memory/GPU topology so the guest driver can online memory correctly.
• Where applicable: Requires GB200-class hardware, so narrow today. Worth the time anyway for the general lesson about fabric-coupled racks — the same assumptions break on any similar architecture.