NVIDIA vs. Nouveau Linux Driver Performance With Extra Kepler Re-Clocking Patches
Last week I posted benchmarks of the AMD proprietary vs. open-source Radeon R600/RadeonSI Gallium3D drivers of various graphics cards on the newest open-source code. Today I'm doing a similar treatment on the NVIDIA GeForce side with seeing how their proprietary driver compares to the latest open-source Nouveau code.
On the NVIDIA binary side was the NVIDIA 358.16 driver release. On the open-source side was Mesa 11.2-devel via the Padoka PPA atop Ubuntu 15.10; this was the latest code available as of Monday. For the kernel DRM side, I was using Karol Herbst's Nouveau DRM branch with the latest Kepler re-clocking patches that aren't yet in the mainline Linux kernel or DRM-Next for that matter. His Nouveau DRM driver was building against Linux 4.3.0 (not 4.4 Git or 4.2 for that matter), so that's where this experimental driver testing was done. Atop this latest Nouveau driver stack, each Kepler graphics card was re-clocked to the highest performance state (0f). Even with this very latest Nouveau code, there isn't yet support for any Kepler automatic re-clocking but still one must boot their system with having the nouveau.pstate=1 module parameter and then write the desired performance state (pstate) to the respective sysfs file.
Testing for this article was done with a NVIDIA GeForce GTX 650, GTX 680, GTX 750 Ti, GTX 760, and GTX 780 Ti. All of these cards are Kepler GPUs with the exception of the original Maxwell GeForce GTX 750 Ti graphics card, which hadn't worked for re-clocking. Additionally, the GTX 750 Ti on Mesa 11.2-devel was only exposing OpenGL 3.3 for now rather than OpenGL 4.1 with the Kepler family. None of the GeForce GTX 900 series could be tested since there still isn't hardware-accelerated open-source OpenGL support until NVIDIA ends up releasing their signed firmware images for these newer Maxwell graphics cards.
All tests for this article were done using the Phoronix Test Suite using both Steam Linux games as well as our classic OpenGL benchmarks.