Owners of aging Radeon hardware running Linux stand to get a smoother experience when their GPU locks up, thanks to fresh open-source work from Valve's graphics team. The effort adds proper IP block soft reset support to the AMDGPU kernel driver, letting the system recover a single hung component without nuking the entire GPU or wiping the contents of VRAM.
The work is led by Timur Kristóf of Valve's open-source Linux GPU driver team, who has spent recent months hauling early Graphics Core Next (GCN) parts forward. That includes moving GCN 1.0 and GCN 1.1 cards onto the modern AMDGPU driver by default instead of the older Radeon driver. His latest patch series reworks how AMDGPU handles resets entirely, deleting defunct code paths and redundant checks before wiring soft reset in as a real GPU recovery method that triggers when a job times out.
Under the old behavior, a hang on much of this hardware meant a full ASIC reset, and on chips like Carrizo and Stoney the existing soft reset code simply failed. The new approach isolates the affected IP block, backs up the contents of all rings tied to it, performs the hardware-specific reset, then restores the rings while minimizing damage to innocent jobs. A GFX block reset, for instance, hits all graphics and compute rings at once. Kristóf also folded in patches from AMD's Alex Deucher originally written for pipe reset, and the team built a Vulkan test case that intentionally hangs the command processor to reliably reproduce the failure.
This first round targets GFX8-class silicon, covering Polaris, Fiji, Tonga, and Carrizo, with fixes for compute rings hanging after a reset and for elevated power draw once the reset completes. Kristóf has said he is also eyeing the same treatment for older GCN hardware, potentially reaching all the way back to GCN 1.0. The changes still need to bake before landing in the mainline AMDGPU driver, but the payoff is a recovery path that keeps your VRAM intact instead of tearing the whole card down.
