Last week Mesa 26.0-devel enabled the ACO back-end by default within the RadeonSI Gallium3D driver for all supported Radeon graphics cards by this open-source Linux driver. This move was done in the name of better performance, faster shader compilation times, and ACO being all-around better than the AMDGPU LLVM back-end these days for both OpenGL and Vulkan use. It was also noted that RadeonSI has “slightly better” viewperf performance with NIR+ACO than using the AMDGPU LLVM back-end. Curious about that SPECViewPerf impact, here are some benchmarks with the recently released AMD Radeon AI PRO R9700 graphics card.
When looking at the ACO performance it’s typically been in the context of desktop graphics/gaming but given the Viewperf mention and recently running lots of tests with the new RDNA4-based Radeon AI PRO R9700 workstation graphics card, I decided to run some benchmarks on the latest Mesa Git code.
Benchmarks with the Radeon AI PRO R9700 were done when using Mesa 26.0-devel with the new ACO default and then repeating the tests using “AMD_DEBUG=usellvm” to fall-back to the AMDGPU LLVM shader compiler back-end. SPECViewPerf was the focus of the runs given Marek’s comments on increased performance with ACO.
Within the ENERGY-O3 and MEDICAL-03 viewsets there was indeed a performance improvement found out of ACO:
With the medical viewset for medical visualization volume rendering techniques and energy fpr the OpendTect seismic visualization application, there were 3~7% better performance using the new ACO default compared to AMDGPU LLVM with the RadeonSI driver.
With the SPECViewPerf benchmarks of the Radeon AI PRO R9700 with the other view-sets, there was no run-time difference observed. But given the competition in the workstation graphics space and the dominance of SPECViewPerf as a benchmark here, a 3~7% improvement is certainly worthwhile on top of their other recent open-source driver optimizations for workstation graphics.
