Linux 7.3 and Intel GPUs: what really changes

Linux 7.3 was released in October 2026. Our servers now ship with this kernel. Here is what it brings on the graphics side, what it doesn’t, and our own measurements on an Intel UHD 770 GPU shared between several VMs.

What 7.3 brings on the graphics side

According to the release notes and Phoronix’s coverage:

Recent Intel GPUs: Xe3P (Nova Lake) graphics are enabled by default, without the force_probe option; Xe3 (Panther Lake) gains a more power-efficient memory bandwidth mode when the display needs little;
video memory management: DMEMCG eviction (per-cgroup GPU memory control) works with the AMD and Intel drivers, and TTM frees memory more aggressively: useful when video memory is full;
outside graphics: faster I/O, better Btrfs performance, groundwork for upcoming AMD processors.

To be clear: none of these new features specifically targets Gen12 GPUs like the UHD 770 (Alder Lake). The Intel changes in 7.3 mostly concern the newest chips.

On our servers, a GPU is shared between several VMs: each VM sees its own GPU device (/dev/dri), with its regular Linux drivers (VAAPI, Vulkan, OpenCL). The i915 driver in the mainline kernel does not support this GPU sharing on these chips: we use an out-of-tree kernel module, rebuilt for each kernel, including 7.3.

Version 2026.10.28 of this module covers kernels 6.18 to 7.3. That is what makes 7.3 usable today, on the host as well as in the VMs.

Measurements: 7.2 vs 7.3

Method

same VM (4 vCPU, like the Perf 2 plan), same GPU share, same GPU-sharing module 2026.10.28; only the kernel changes: 7.2.11 (Debian backports) then 7.3.0;
each test is run 3 times, and we keep the median value;
the GPU is not used by other VMs during the measurement;
source video: 4K HEVC 10-bit HDR, 24 frames/s;
WebGL: our public benchmark webgl-bench.html (full-screen 1080p shader, 10 draws per frame, one pixel read back per frame to force rendering to finish);
OpenCL: clpeak (single-precision compute and memory bandwidth).

Results

Test	Linux 7.2.11	Linux 7.3.0
4K HEVC 10-bit → 1080p H.264 (VAAPI, low-power)	12.1× real time	12.1× real time
4K HEVC 10-bit → 1080p HEVC (VAAPI, low-power)	11.9×	11.9×
4K HEVC 10-bit decode only	13.4×	13.1×
WebGL 1080p (Chromium, Vulkan)	281 frames/s	279 frames/s
OpenCL, single-precision compute	737.6 GFLOPS	737.6 GFLOPS
OpenCL, memory bandwidth	42.0 GB/s	42.0 GB/s

GPU performance is identical: the differences are within the margin of measurement. This is no surprise: the work is done by the GPU and its firmware (GuC/HuC), not by the kernel.

Stability

On an identical test server, with the host on 7.3: several VMs ran simultaneously for 20 minutes, each with its own workload (Chrome in 3D, 4K HDR transcoding with tone mapping, 4K AV1, a small OpenVINO AI model). Result: no GPU hang in the VMs, no reset on the host.

What matters for stability is not the kernel version: it is that the host and the VM use the same compute engine (CCS) configuration of the GPU. Our VM images are configured accordingly.

So why move to 7.3?

Same GPU performance, no regression measured;
the general kernel gains: I/O, Btrfs, security fixes, recent hardware;
an up-to-date kernel for the years to come, with a GPU-sharing module maintained for this version.

In other words: 7.3 doesn’t make your GPU faster, but you can adopt it without losing anything.

Linux GPU is 7.3-ready

New Linux GPU servers ship with Linux 7.3, the GPU-sharing module 2026.10.28, and the VAAPI, Vulkan, OpenCL and OpenVINO drivers already installed. A single command, gpu-check, verifies that everything works from the first login. From €14.99 excl. VAT/month, root access, dedicated IPv6.

Sources: Phoronix, Linux 7.3 released, Phoronix, Linux 7.3 graphics changes.