AMD Working On More Precise GPU Reset Capabilities For Their Linux Driver

While the hope remains that GPU resets are a very infrequent task, AMD Linux driver engineers have recently been working on the ability to support a per-queue GC reset capability for more precise reset capabilities when needed.

With the latest patches for the AMDGPU kernel driver code, the driver would be able to reset the graphics/compute ring where an error occurs and to just lose the jobs pending in that queue as opposed to carrying out a full GPU reset. But if the per-queue/ring reset doesn't work, the AMDGPU driver would continue to perform a full GPU reset as needed.

An updated set of 53 patches were posted today for introducing reset capabilities on a per-ring basis in the AMDGPU driver. This is then wired up across the recent generations of AMD graphics processors, including the upcoming GFX12 (RDNA4) hardware.

The hope is that this per-ring reset capability is more pleasant when encountering a specific job that's hung rather than going the route of a full "legacy" GPU reset.

Those interested in learning more can do so via this patch series now under review for GC per-queue resets.