News › performance
By The Linux Camp News, October 30, 2026
PerfOpt lets supported AMD integrated GPUs bypass the IOMMU for system-memory access. The merge limits eligibility to GPUs in identity domains, rather than all Radeon hardware or IOMMU configurations.
Linus Torvalds merged the IOMMU updates containing AMD PerfOpt for Linux 7.4 on October 23, 2026, according to the commit’s -1000 timestamp. The optimization targets AMD integrated graphics through the AMDGPU driver. The sources retrieved October 30 establish integration, not a final Linux 7.4 release or distribution availability. 2 1
Phoronix describes the change as letting an integrated GPU bypass the IOMMU, the device memory-management unit, when directly accessing system memory. PerfOpt therefore changes the memory-access path by allowing that stage to be skipped, rather than adding GPU compute capabilities. The maintainer’s merge summary confines activation to supported integrated GPUs in identity domains, meaning one-to-one address mappings. 1 2 3
Phoronix reports that PerfOpt is enabled by default with AMDGPU on Radeon integrated graphics. That needs to be read alongside the maintainer’s narrower eligibility statement: neither every Radeon GPU nor every IOMMU configuration is covered. The supplied summary does not identify supported GPU generations or explain behavior outside identity domains, so it is insufficient grounds for changing an operator’s IOMMU configuration. 1 2
The workloads worth evaluating are integrated-GPU jobs, particularly AI: Phoronix attributes performance improvements to PerfOpt in its benchmarks. This is a benchmark publisher’s assessment, not a generally established speedup. The supplied article references earlier testing but provides no benchmark tables, test configurations or variability from which to estimate the gain on another machine. 1
When evaluating a running kernel containing the change, inspect its dmesg output for “PerfOpt,” as Phoronix advises, to verify that AMDGPU armed the optimization. That activation check matters because support is conditional: an installed kernel package, or even a running version number, is not confirmation that this GPU enabled PerfOpt. The log check establishes activation, not the size of any workload improvement. 1 2