nvgd

nvgd is an independent, Mesa-based graphics driver focused on NVIDIA Turing and newer GPUs (GTX 16xx, RTX 20xx/30xx/40xx). OpenGL and Vulkan run natively with no translation layer: OpenGL 4.6 through the nvc0 Gallium driver with the NAK shader compiler, and Vulkan 1.4 through NVK. Nothing goes through Zink or ANGLE.

The kernel side is nvgd-dkms, a nouveau fork with native GSP-RM 615 firmware support. Arch/CachyOS packages live in nvgd-repo.

This is an independent project. It is not affiliated with Mesa, nouveau or NVIDIA and is not submitted to upstream Mesa. The upstream branch tracks official Mesa and is used as the rebase base.

Performance: nvgd vs. the proprietary driver (RTX 3050)

Same machine and GNOME/Wayland desktop. The OpenGL rows below are the existing glmark2 2023.01 measurements at its default 800x600 resolution.

Mode	nvgd	NVIDIA 615.78.08	nvgd / NVIDIA
glmark2 --off-screen (driver only, no presentation)	35,236	31,803	1.11x
glmark2-wayland (windowed, native Wayland)	9,902	2,653	3.73x
glmark2 (windowed, X11/GLX via Xwayland)	4,984	4,016	1.24x
vkmark (Vulkan, native Wayland, 800x600, immediate)	14,076	12,925	1.09x
vkmark (Vulkan, NVK re-measurement, 2026-10-23)	15,510	not re-measured	—

The Vulkan result is the median of three complete vkmark 2025.01 runs: 14,076 / 13,401 / 14,090, measured on 2026-10-22 (UTC-03). NVIDIA 615.78.08 scored 13,399 / 12,925 / 12,792 (median 12,925) after reboot, with identical benchmark binary/data hashes and settings. NVK's aggregate score is 8.9% higher in this workload, not in every scene: NVIDIA leads in 8 of the 13 scene medians. The kernel differs (nvgd: 7.3; NVIDIA: 6.18 LTS), and these are normal desktop sessions, not an isolated driver-only comparison. Scores from vkmark and glmark2 are not comparable across benchmarks, and neither score measures input latency. Vulkan settings, raw logs and the matching NVIDIA procedure.

A fresh NVK-side re-measurement on 2026-10-23, with the same vkmark binary, plugin and data hashes, scored 15,510 / 15,549 / 15,497 (median 15,510). This is not a new paired result: NVIDIA was not re-run under these conditions, so the 1.09x comparison above remains the reference. The higher NVK score reflects the current system state, not a driver-code change. Re-measurement logs and environment.

Setup:

• GPU: GeForce RTX 3050 6 GB (GA107), PCIe 3.0 x8
• CPU: Intel Xeon E5-2697 v3
• OS: CachyOS, GNOME (Wayland)
• nvgd OpenGL rows: kernel 7.2.11 + nvgd-dkms, nvgd Mesa
• nvgd Vulkan row: kernel 7.3.0-3-cachyos, nvgd-dkms 1.0.0-20, Mesa a1c7b37fcb7b, vkmark 1:2025.01-2.1; fixed B8G8R8A8Unorm format, 13 default scenes at 10 seconds each, no overlay/validation layers
• Proprietary: kernel 6.18 LTS + nvidia-open 615.78.08

Per scene (--off-screen)

nvgd wins 22 of the 33 scenes.

Scene	nvgd	NVIDIA 615.78.08	nvgd / NVIDIA
build use-vbo=false	4887	4834	1.01x
build use-vbo=true	67300	53369	1.26x
texture texture-filter=nearest	51382	56129	0.92x
texture texture-filter=linear	50478	54459	0.93x
texture texture-filter=mipmap	55082	56177	0.98x
shading shading=gouraud	50654	41633	1.22x
shading shading=blinn-phong-inf	49386	40908	1.21x
shading shading=phong	45700	37598	1.22x
shading shading=cel	46051	37349	1.23x
bump bump-render=high-poly	25109	22499	1.12x
bump bump-render=normals	73438	60836	1.21x
bump bump-render=height	67400	59266	1.14x
effect2d kernel=0,1,0;1,-4,1;0,1,0;	32915	40909	0.80x
effect2d kernel=1,1,1,1,1;1,1,1,1,1;1,1,1,1,1;	20601	22464	0.92x
pulsar light=false:quads=5:texture=false	45799	41909	1.09x
desktop blur-radius=5:effect=blur:passes=1:separable=...	7563	4542	1.67x
desktop effect=shadow:windows=4	13201	7998	1.65x
buffer columns=200:interleave=false:update-dispersio...	792	946	0.84x
buffer columns=200:interleave=false:update-dispersio...	726	918	0.79x
buffer columns=200:interleave=true:update-dispersion...	539	1057	0.51x
ideas speed=duration	9204	10784	0.85x
jellyfish	21271	19205	1.11x
terrain	1148	1327	0.87x
shadow	24634	12181	2.02x
refract	2846	2865	0.99x
conditionals fragment-steps=0:vertex-steps=0	50868	45516	1.12x
conditionals fragment-steps=5:vertex-steps=0	49038	44381	1.10x
conditionals fragment-steps=0:vertex-steps=5	50229	44806	1.12x
function fragment-complexity=low:fragment-steps=5	49961	44971	1.11x
function fragment-complexity=medium:fragment-steps=5	47445	43534	1.09x
loop fragment-loop=false:fragment-steps=5:vertex-s...	49247	45091	1.09x
loop fragment-steps=5:fragment-uniform=false:verte...	49221	44681	1.10x
loop fragment-steps=5:fragment-uniform=true:vertex...	48726	44417	1.10x

Where the proprietary driver still leads:

• Very light scenes (texture, effect2d) are bound by the round trip to the GPU. NVIDIA's driver submits work straight from userspace through the doorbell, while nvgd still goes through a kernel ioctl every frame (~4.5 µs on this CPU).
• ideas is bound by per-draw CPU cost in Mesa.
• terrain and refract are GPU-bound, so shader code quality from the compiler matters most.

To reproduce:

glmark2 --off-screen
glmark2-wayland
glmark2

Vulkan (three full runs, with a three-second cube warm-up before each):

python3 docs/benchmarks/run-vkmark.py nvgd /path/to/new-nvgd-results
# After rebooting into NVIDIA, from a Wayland desktop:
python3 docs/benchmarks/run-vkmark.py nvidia /path/to/new-nvidia-results \
  --reference docs/benchmarks/vulkan-2026-10-22/nvgd/environment.json

What nvgd changes compared to Mesa

• NAK for OpenGL: NVK's Rust compiler compiles the nvc0 driver's shaders on Turing+. Anything NAK does not cover yet (images, SSBOs, legacy compatibility-profile varyings) falls back to nv50_ir.
• Deferred submission: with a single context, work is submitted at flush/fence time instead of after every draw. It is still submitted early once the GPU goes idle, and immediately when it touches shared dma-buf buffers.
• Fences: fence waits poll the coherent fence page before sleeping in the kernel, and fence BOs are recycled.
• Colour compression on private render targets, with lossless reduce-color thresholds.
• Buffers: idle VRAM buffers are mapped directly for write-only maps.
• State validation: fixed duplicated texture/UBO references that grew without bound.
• Rendering fixes for Chromium/Electron (Discord, Brave), GLX and GPU Screen Recorder.
• Native video: H.264, HEVC and AV1 VA-API decode submit directly to NVDEC through Gallium/DRM, and VP9 through a full spec header parser with backward probability adaptation, without loading NVK or translating to Vulkan Video. H.264 encode uses NVENC directly, including separate Gallium input planes. Architecture, validated limits and codec roadmap.

Environment variables

Variable	Default	Effect
NVGD_USE_NAK	true	use NAK for OpenGL shaders
NVGD_DEFER_KICK	true	deferred submission
NVGD_KICK_IDLE_MIN / NVGD_KICK_BATCH	96 / 256	early submission when the GPU is idle / batch cap
NVGD_FENCE_SPIN_US	200	polling time before the kernel fence wait
NVGD_COLOR_COMPRESSION	true	colour compression on Turing+
NVGD_DIRECT_VRAM_MAP	true	direct maps of idle VRAM buffers
NVGD_DEBUG_NAK	false	log which compiler each shader used

Building

Same as Mesa (see README.mesa.rst and https://docs.mesa3d.org/install.html):

meson setup build -Dgallium-drivers=nouveau,zink,softpipe -Dvulkan-drivers=nouveau
ninja -C build

Contributing

Pull requests are welcome. Open them against main; every PR is reviewed and accepted or declined by the maintainer. For performance changes, include glmark2 numbers before and after plus the results of the rendering tests. Use Issues for bugs and ideas.

License

Same licenses as Mesa (mostly MIT; see docs/license.rst).