With Mesa Git and Linux 4.9 I have done AMDGPU vs. Radeon, AMD OpenGL vs. Vulkan benchmarks, and other tests while in this article is our first side-by-side comparison of AMD Radeon GPUs on their open-source driver against NVIDIA GPUs on Nouveau when using Linux 4.9 Git and Mesa Git as of this week.
For your viewing pleasure this Friday are some open-source AMD vs. NVIDIA numbers when using the latest open-source code on each side. Linux 4.9-rc1 was used while Ubuntu 16.10 paired with the Padoka PPA led to Mesa Git as of earlier this week plus LLVM 4.0 SVN. As covered recently, there are no Nouveau driver changes for Linux 4.9 while we had hoped the boost patches would land. Thus the re-clocking is still quite poor for this open-source NVIDIA driver stack. For the Nouveau tests I manually re-clocked each graphics card to the highest performance state (0f) after first re-clocking the cards to the 0a performance state for helping some of the GPUs that otherwise fail with memory re-clocking at 0f, as Nouveau developers have expressed this is the preferred approach for testing.
Tested on the red side for this article was the HD 7950, R9 285, RX 460, and RX 480. Unfortunately the R9 Fury was in a regressed state on Linux 4.9-rc1 and my R9 290 also is still in a messed up state on the latest open-source driver stack. On the green side was the GTX 650, GTX 680, GTX 760, and GTX 780 Ti re-clocked as far as they could go with the latest mainline kernel. Only NVIDIA Kepler cards were used since they offer the best (but still limited) re-clocking support, there isn't any re-clocking support for newer GTX 900 Maxwell cards on the open-source driver, and there isn't yet any open-source 3D acceleration for GTX 1000 Pascal cards. Of course, under the NVIDIA binary driver you'll find the best GeForce performance on Linux.
Let's now look at these open-source 3D performance results for various OpenGL workloads on this Mesa Git / Linux 4.9 roundabout using the Phoronix Test Suite.