AMD GPU Driver Regression Slows AI Workloads 43x on Linux

Ubuntu engineers have issued a warning that the upcoming Linux kernel 7.0.0-28.28 causes a severe performance regression in the AMDGPU driver. This update creates a bottleneck for users running compute-heavy applications on Linux, potentially making routine tasks significantly slower. Buyers relying on AMD GPUs for machine learning or AI inference should hold off on updating their systems until Canonical releases a fix.

Canonical advises avoiding kernel 7.0.0-28.28 for compute users

The regression specifically targets ROCm-based applications, which are essential for running AI models like Stable Diffusion XL (SDXL) through interfaces such as ComfyUI. Canonical advises that users depending on AMD ROCm compute capabilities should avoid installing the 7.0.0-28.28 kernel update. The issue stems from changes in the upstream Linux 7.0 stable series, meaning it impacts other distributions like CachyOS and Fedora as well.

The performance drop is quantified as a 43 times reduction in throughput. Workloads that previously completed in approximately nine seconds are now taking up to 388 seconds after the upgrade. This massive slowdown occurs because the kernel update introduces a bottleneck in the AMDGPU driver's handling of these specific compute tasks.

While the impact on AI inference is severe, typical gaming workloads are unlikely to suffer significant performance losses. The regression is isolated to specific compute-heavy ROCm operations rather than general graphics rendering. Users playing standard games on Linux should not experience the same dramatic slowdowns reported by AI developers.

Canonical recommends that users avoid the update until a fix is available. The company is working to resolve the upstream issue in the Linux 7.0 stable series. Until then, AMD GPU users on Ubuntu, CachyOS, and Fedora should stick to their current kernel versions to maintain performance stability.