RadeonSI ACO vs. LLVM Backends For AMD Strix Halo

With the RadeonSI Gallium3D driver now defaulting to the ACO compiler back-end for all Radeon GPUs rather than the conventional AMDGPU LLVM shader compiler back-end, I ran some quick comparison benchmarks on AMD Ryzen AI Max+ “Strix Halo” with Radeon 8060S Graphics for comparison.

Over the weekend I ran some quick OpenGL benchmarks on AMD Strix Halo for comparing the performance impact of Mesa 26.0-devel with its new ACO default for all GPUs compared to using AMDGPU LLVM as the prior default and can still be enabled via the “AMD_DEBUG=usellvm” environment variable override.

For this quick round of OpenGL ACO vs. LLVM benchmarking I used the popular Framework Desktop platform.

Mesa 26.0-devel from the Mesa ACO PPA was used for testing on Ubuntu 25.10 with the Linux 6.17 kernel. The only change between runs was falling back to the AMDGPU LLVM back-end for comparison.

The new SuperTuxKart 1.5 was showing some minor enhancements when using RadeonSI with ACO for the RDNA 3.5 integrated graphics.

And in the more demanding OpenGL benchmarks too like GravityMark.

Or the Unigine benchmarks also performed slightly better with this new ACO default.

For the most part there were some nice incremental gains for RadeonSI on Mesa 26.0-devel with its new ACO compiler back-end default compared to RadeonSI. No issues were encountered on Strix Halo with this ACO default. The performance was already in great shape with the AMDGPU LLVM back-end while now is slightly better with ACO and more pronounced for faster shader compilation / speedier game load times. Granted, most Linux games in 2025+ are ultimately going through the Vulkan API these days with RADV where ACO has long been the default, but for any retro gamers or other heavy OpenGL users, this ACO default is nice to see happen.