Mesa’s open-source NVK Vulkan driver for NVIDIA GPUs is on a bit of a tear lately. After a recent merge that adds compression support—a change expected to reduce bandwidth pressure and improve performance—another optimization landed earlier in the week that should be particularly helpful for GeForce RTX 20-series “Turing” cards.
The catch: the compression work is tied to newer plumbing in the stack, and the performance upside will ultimately depend on your kernel + Mesa combo.
NVK compression support lands, but benchmarks still pending
This week’s headline NVK update is the merge of compression-related support in Mesa. The goal is straightforward: let NVK take advantage of hardware compression paths (including depth/stencil “zeta” compression) and related memory management improvements to cut down on unnecessary memory traffic.
That’s the kind of change that can translate into real-world gains—but there aren’t public benchmark numbers attached to the merge itself yet, so it’s best framed as “promising” rather than “proven.”
One important detail for Linux users: the work leans on new Nouveau kernel functionality, reported as coming in Linux 6.19, meaning you’ll need a sufficiently new kernel (plus a recent Mesa) to see the full benefit.
A second NVK optimization targets Turing performance
Separate from compression, NVK’s NAK compiler also picked up an optimization aimed at better matching Turing hardware behavior.
The change—described as using the hardware’s max warps_per_sm value—should ensure NVK propagates the correct warps-per-SM values across different Turing GPUs. In practical terms, that may help workloads that are heavy on compute (including some modern games and GPU-driven effects), where scheduler and occupancy decisions can matter.
Again, no performance figures were included alongside the merge, so this falls into the “should help, but we’ll need benchmarks” category.
The bigger picture: Nouveau/NVK is moving fast
Between compression support and compiler-level tuning, NVK continues to stack up meaningful improvements at a steady clip. If you’ve been following open-source NVIDIA driver progress, the next round of real-world testing should be interesting—especially on older architectures like Turing where small efficiency wins can add up.
If you want context from recent comparisons, there have also been fresh benchmark roundups spanning older GeForce cards through the latest GPUs, plus broader AMD vs. NVIDIA open-source performance looks for 2025.