nestrilabs/virtio-nvgpu

A virtio device for sharing one NVIDIA GPU across several VMs, at bare-metal speed.

virtio-nvgpu – Share One NVIDIA GPU Across Multiple VMs

What it does

  • Provides a virtio device that lets several Linux virtual machines (VMs) use the same physical NVIDIA GPU without needing a vGPU license or PCI‑passthrough.
  • The host runs the official NVIDIA driver; the guest kernels load a small driver module that forwards the actual NVIDIA driver calls to the host. GPU commands are written directly into GPU memory, so the workload runs at near‑bare‑metal speed.

Why it matters

  • Traditional PCI‑passthrough gives a whole GPU to one VM, limiting consolidation.
  • NVIDIA’s vGPU solution is expensive and limited to datacenter‑class cards.
  • Existing API‑forwarding solutions (Venus, virgl) are slow and cannot expose NVENC or CUDA.
  • virtio‑nvgpu sidesteps these limits by sharing the driver itself, keeping performance high while allowing many guests.

Performance claims

  • On an RTX 3060, frame‑rendering latency is within 2 % of native bare‑metal performance for typical game frames (>2 ms).
  • CPU overhead is comparable to running on the host directly.
  • The authors demonstrated 12 VMs on a single card, each encoding 720p60 H.264 without dropped frames (limited by VRAM, host resources, and the 12‑session NVENC cap on GeForce cards).

Typical use cases

  • Cloud‑gaming or remote‑desktop services that need real‑time GPU acceleration.
  • CI pipelines or server‑less AI workloads that require a physical GPU but want isolation via VMs.
  • Any headless workload that benefits from CUDA, Vulkan, or NVENC on consumer‑grade NVIDIA GPUs.

Limitations & requirements

  • Only Linux hosts and guests are supported; the driver version must be 535.129.03 or newer.
  • Security is limited to the host driver – it is not a hardware isolation boundary. Untrusted tenants should use passthrough or NVIDIA’s official vGPU instead.
  • The project is early‑stage: tested on two GPU models, CUDA works but lacks extensive benchmarking.

How it is organized

  • driver/ – Guest‑side kernel module (GPL‑2.0).
  • device/ – VMM‑agnostic Rust implementation of the virtio device (Apache‑2.0).
  • protocol/ – Shared protocol definitions (BSD‑3/GPL‑2.0).

Further reading

  • Architecture details: ARCHITECTURE.md
  • Benchmark methodology and results: BENCHMARKS.md
  • Security considerations: SECURITY.md

Bottom line virtio‑nvgpu is a niche but practical solution for sharing a single consumer‑grade NVIDIA GPU across many Linux VMs while retaining near‑native performance, aimed at developers and operators who need GPU‑accelerated workloads in a multi‑tenant environment without the cost of NVIDIA’s vGPU licensing.

Related

  • Dispatch
    virtio-nvgpu enables near‑native NVIDIA GPU performance in KVM guests

    virtio-nvgpu forwards NVIDIA driver ioctls at the ABI level, letting Linux KVM guests render and encode with less than 2 % overhead compared to bare metal, while sharing a single GPU across multiple VMs.

  • Project
    NVIDIA/gds-nvidia-fs

    A kernel driver (`nvidia-fs.ko`) that implements NVIDIA GPUDirect Storage, allowing supported NVMe, computational‑storage, or RDMA filesystems to move data directly into GPU memory, boosting AI/ML data‑pipeline throughput.

  • Project
    NVIDIA/k8s-device-plugin

    The NVIDIA k8s‑device‑plugin is an official DaemonSet that implements Kubernetes’ device‑plugin API for NVIDIA GPUs. It discovers GPUs, reports health, and makes them schedulable via the `nvidia.com/gpu` resource (with optional MIG and shared‑GPU modes). Install the NVIDIA driver and container toolkit on each node, deploy the plugin (static manifest or Helm), then request GPUs in pod specs. Configuration is possible via flags, env‑vars, or a YAML file, covering MIG strategy, device‑list delivery, shared‑GPU (time‑slicing or MPS), and more. The plugin is beta, lacks full health‑check/cleanup, and only the official NVIDIA version is supported.

  • Dispatch
    vLLM Adds NVIDIA Vera Rubin NVL72 Support with 7.8× Throughput Gains

    vLLM now runs on NVIDIA Vera Rubin NVL72 GPUs, delivering up to 7.8× higher inference throughput than GB200 NVL72 and showing strong early performance on AgentX and MLPerf benchmarks.

  • Project
    CVCUDA/CV-CUDA

    CV‑CUDA is NVIDIA’s open‑source GPU‑accelerated computer‑vision library. It provides fast, scalable image/video operators (resize, filter, color conversion, etc.) that run entirely on the GPU and integrate with C/C++ and Python AI frameworks. Install via pre‑built wheels (`cvcuda-cu12` or `cvcuda-cu13`) for Linux, decode images directly to GPU with `nvimgcodec`, and chain CV‑CUDA operators before feeding data to deep‑learning models. The library targets modern NVIDIA GPUs (Turing and newer) and is used in production systems such as Microsoft Bing visual search and Tencent Cloud video processing.