NVIDIA’s latest flagship GPUs—the GeForce RTX 5090 and the RTX PRO 6000—have reportedly run into a serious virtualization bug that causes the cards to become unresponsive after extended use in VM environments.
The issue first surfaced through CloudRift, a GPU cloud service for developers. According to their findings, after just a few days of VM assignment, affected GPUs stop responding altogether. The only way to recover them is to reboot the entire host node, a process that disrupts multiple client environments.
Which GPUs Are Affected?
Currently, the virtualization bug appears to be isolated to Blackwell-based RTX 5090 and RTX PRO 6000 models. NVIDIA’s older RTX 4090, Hopper H100, and B200 Blackwell GPUs are not experiencing the same issue.
The problem arises when the GPU is passed through a VM using the VFIO device driver. After a Function Level Reset (FLR), the card fails to respond, creating a kernel soft lock that freezes both host and client systems.
Community & Industry Reports
CloudRift isn’t alone in reporting problems. A Proxmox user also described a complete host crash after shutting down a Windows client VM with an RTX 5090 assigned.
NVIDIA has reportedly acknowledged the issue, stating they were able to reproduce the bug internally and are investigating a solution.
Bug Bounty and Potential Fixes
Due to the severity of the issue, CloudRift has offered a $1,000 bug bounty to anyone who can deliver a fix or workaround. Considering how crucial these GPUs are for AI workloads and virtualization environments, industry watchers expect NVIDIA to release a patch or driver update soon.
