Research · Virtualization

GPU virtualization and passthrough research

How we run GPUs inside virtual machines in production, what the settings are worth, and what broke along the way.

CloudRift, which operates RiftStack, runs three GPU virtualization strategies in production: VFIO passthrough, NVIDIA MIG with AI Enterprise vGPU drivers, and AMD SR-IOV with ROCm. The overview article covers the host-side mechanics, the driver lifecycle, the libvirt domain XML and the trade-offs of each. The host setup guide goes deeper on QEMU/KVM passthrough on Linux: IOMMU groups, driver binding, and the resets and boot failures that cost weekends.

The other two articles come out of running this at scale. One benchmarks six VM configurations to find when vCPU pinning and guest NUMA topology help, when they make no measurable difference, and when they slow multi-GPU jobs. The other is the reset bug we hit on the RTX 5090 and RTX PRO 6000 under passthrough, where a function-level reset times out and only a host reboot brings the card back. We posted a $1,000 bounty, the hardware press picked it up, and a community workaround resolved it. NVIDIA has since acknowledged the bug.