By ByteBot
Incus 7.3 just solved one of the more persistent headaches in Linux virtualization: GPU sharing. Until now, one GPU meant one VM. You picked a winner; everyone else ran software rendering. The July 31 release changes that with a native-context GPU type that lets multiple virtual machines share a single physical GPU at the same time — no SR-IOV hardware requirement, no enterprise licensing bill.
What Actually Changed
Traditional GPU passthrough assigns an entire physical GPU exclusively to one VM through VFIO/IOMMU. The VM sees real hardware and gets near-native performance — but the GPU is locked to that VM alone. If you need GPU acceleration in three VMs and have one card, you were stuck buying extra hardware, dealing with NVIDIA’s vGPU licensing, or swapping the GPU between VMs with downtime.
Incus 7.3 introduces a new device type called native-context that pairs virtio-gpu with virglrenderer. The key breakthrough is where the GPU sharing happens: DRM native context remotes at the kernel UAPI layer, not the application API layer. Each guest process gets its own GPU address space on the host. The result is that multiple VMs can be served from one physical GPU simultaneously, with performance that benchmarks within a few percent of bare metal.
What You Need to Use It
This is not plug-and-play for everyone. Here is what the host requires:
- QEMU 11.0.0 or newer — the native-context support lives in QEMU’s virtio-gpu implementation
- virglrenderer 1.0.0+ with DRM native context enabled — your distro’s package may or may not be built with this flag
- A supported GPU: AMD (AMDGPU, usable since virglrenderer ~1.1), Intel i915 (usable since ~1.3), or Qualcomm Freedreno (fully upstream). Support varies by virglrenderer version.
- Matching guest drivers — the guest needs a virtio-gpu DRM driver that supports native contexts
That last point about supported GPUs deserves emphasis: NVIDIA is not on the list. If your homelab or workstation runs NVIDIA for CUDA-based ML workloads, this feature does not help you. For NVIDIA GPU compute in VMs, you still need passthrough (one VM) or vGPU licensing. This is an AMD and Intel story — at least for now.
Upgrade Now: 13 Security Fixes, 9 Critical
The GPU feature gets the headline, but the security fixes are the actual reason to upgrade immediately. Incus 7.3 patches 13 vulnerabilities — 9 rated critical, 2 high, 2 medium. The attack surfaces include arbitrary file read and write on the host via crafted images, path traversal, symbolic link attacks, and bypasses of project isolation. If you run Incus in any shared or production environment, waiting on this upgrade is not an option.
Smaller Wins Worth Knowing
Two quality-of-life improvements round out the release. The new incus port-forward command creates a local TCP listener that tunnels directly to a port inside a container or VM — no manual network bridge configuration. And incus low-level nvram adds direct CLI access to UEFI NVRAM variables, letting you change boot order and Secure Boot settings without opening the guest firmware interface. Both matter if you are building automated provisioning pipelines or doing active development inside Incus VMs. The virglrenderer project has been steadily expanding support across GPU families, and Incus 7.3 brings that work to a practical shipping state.
The Honest Take
Incus 7.3’s native-context GPU sharing is a real advancement for AMD and Intel GPU users who need more than one VM to have graphics acceleration. The gap it closes — between expensive SR-IOV enterprise hardware and VFIO’s one-GPU-one-VM constraint — is real. For homelab setups, developer workstations, and open-source cloud infrastructure built on AMD or Intel GPUs, this changes what is practical without proprietary tooling.
NVIDIA GPU users running CUDA workloads in VMs still have limited options. That is not Incus’s fault — it reflects where the upstream GPU vendor has invested in open virtualization support. The virglrenderer community continues working on improved NVIDIA support, but it is not there in any practical sense yet.
If you are already running Incus, update to 7.3 now for the security fixes regardless. If you have been evaluating whether Incus fits your GPU-heavy VM workflows and you are running AMD or Intel hardware, the 7.3 release makes the case stronger.
