AMD GPU hangs and resets under load: 'ring gfx_0.0.0 timeout'
The diagnosis and the triage order (evidence, then update, then RADV, then overclock, only then kernel params) are excellent, and most specifics verify exactly. The 2000 ms watchdog is right: amdgpu_drv.c:365 reads "GPU lockup timeout in ms (default: 2000...), format: [single value for all] or [GFX,Compute,SDMA,Video]". The RADV_DEBUG strings are verbatim from docs.mesa3d.org/envvars.html, including hang's "$HOME/radv_dumps_<pid>_<time>". noretry's "(0 = retry enabled, 1 = retry disabled, -1 auto (default))" is verbatim from amdgpu_drv.c:715. The devcoredump path is valid: devcoredump.c:421 creates a devcoredump symlink on the failing device, so /sys/class/drm/card1/device/devcoredump/data resolves. Two factual errors. (1) The amdgpu.gpu_recovery=1 bullet is wrong, and wrong in the direction that wastes the reader's reboot: amdgpu_device_should_recover_gpu() (amdgpu_device.c:4878) returns true under the -1 auto default for everything except a short legacy list (SI, CIK, Carrizo, Stoney, Cyan Skillfish). Recovery is NOT "disabled outside SR-IOV". On the RDNA2/RDNA3 hardware this record targets, auto already enables it and gpu_recovery=1 is a no-op. (2) The devcoredump does not survive until "the next boot": include/linux/devcoredump.h:16 is #define DEVCD_TIMEOUT (HZ * 60 * 5) with the comment "if data isn't read by userspace after 5 minutes then delete it". A reader who thinks they can grab it tomorrow will find it gone. Also, the sudo tee in step 5 truncates zz-local.conf, silently discarding any parameter the reader added from a sibling record.
gpu_recovery is declared with module_param_named_unsafe in the kernel: setting it taints the kernel, and a *failed* reset can wedge the machine harder than the original hang would have, so save your work before testing. Setting lockup_timeout too high converts a recoverable ring hang into a multi-second (or permanent) full-desktop freeze. Disabling EXPO/DOCP drops your RAM to JEDEC base speed until you re-enable it. Never leave RADV_DEBUG=hang on permanently, because it forces synchronisation and costs significant performance.
Symptom
Mid-game (or during a shader compile, or in a VR session) the screen freezes for a few seconds, the game crashes or the whole session dies back to a TTY, and journalctl -k -b -1 shows:
amdgpu 0000:03:00.0: amdgpu: ring gfx_0.0.0 timeout, signaled seq=7073226, emitted seq=7073228
amdgpu 0000:03:00.0: amdgpu: Process information: process gamename pid 4711 thread gamename pid 4711
amdgpu 0000:03:00.0: amdgpu: GPU reset begin!
...
amdgpu 0000:03:00.0: amdgpu: GPU reset(1) succeeded!
When the reset does not work you get GPU reset(1) failed with error -110 and a hard lock: no SysRq, only the power button. The milder variant just logs ring gfx_0.0.0 timeout, but soft recovered and the game stutters.
Cause
A command submitted to the graphics ring did not signal its completion fence within the scheduler's watchdog window (2000 ms by default, per amdgpu's own lockup_timeout module parameter description), so the driver declared the ring hung and attempted a reset. That is a *symptom*, not a diagnosis. The actual culprit is almost always one of three things, in rough order of frequency on RDNA2/RDNA3 desktops:
- **An unstable overclock or undervolt.** DDR5 EXPO/DOCP profiles, PBO/Curve Optimizer, and LACT/CoreCtrl undervolt profiles are the single most common cause of gfx ring timeouts on desktop Radeon cards.
- **A Mesa/RADV shader bug.** A specific shader compiles to something the hardware chokes on. Reproduces with one game and never with another.
- **A genuine driver/firmware bug** for that ASIC and kernel combination, or marginal power delivery.
Note that GPU reset is already *enabled* by default on this hardware: amdgpu's gpu_recovery defaults to -1 (auto), and amdgpu_device_should_recover_gpu() returns true under auto for every ASIC except a legacy list (SI, CIK, Carrizo, Stoney, Cyan Skillfish). So on RDNA2/RDNA3 the "GPU reset begin!" you are seeing IS the auto path working. A failed reset is a firmware/hardware problem, not a missing kernel parameter.
This is a different failure from flip_done timed out (display pipeline, fixed with amdgpu.dcdebugmask) and from idle GFXOFF lockups. Those hang while doing nothing, this one hangs under load.
Fix
**1. Collect the evidence while it is still there: the crash dump self-deletes after FIVE MINUTES.**
The kernel's devcoredump has a hard 5-minute expiry (DEVCD_TIMEOUT (HZ * 60 * 5)), so grab it in the same session as the hang, not after a reboot:
journalctl -k -b -1 | grep -iE 'amdgpu|drm|ring .* timeout' | tail -60
lspci -nnk -d ::03xx
uname -r; pacman -Q mesa vulkan-radeon linux linux-firmware
ls /sys/class/drm/card*/device/devcoredump/ 2>/dev/null
sudo cat /sys/class/drm/card1/device/devcoredump/data > ~/amdgpu-coredump.txt
**2. Update first.** Ring timeouts are fixed upstream constantly. A stale Mesa or kernel is the cheapest thing to rule out.
omarchy update # Omarchy 4
sudo pacman -Syu # plain Arch / EndeavourOS / CachyOS
**3. Decide whether it is userspace or hardware.** Add to the game's Steam launch options, one at a time:
RADV_DEBUG=hang %command% # writes a report to ~/radv_dumps_<pid>_<time>
RADV_DEBUG=nongg %command% # disables NGG on GFX10/10.3
RADV_DEBUG=zerovram %command% # zero-init VRAM allocations
If the hang stops under hang or nongg, it is a Mesa bug. File the dump at gitlab.freedesktop.org/mesa/mesa and stay on the workaround. If it hangs identically under all of them, go to step 4.
**4. Take the overclock out of the picture.** Reboot into firmware setup and disable EXPO/DOCP (RAM back to JEDEC) and PBO/Curve Optimizer, and remove any GPU undervolt:
systemctl status lactd 2>/dev/null && sudo systemctl stop lactd
ls ~/.config/corectrl/ 2>/dev/null
cat /sys/class/drm/card1/device/pp_od_clk_voltage 2>/dev/null
Run the game for an hour. If it is now stable, reintroduce one setting at a time.
**5. Only then reach for kernel parameters.** On Omarchy 4 add them as a drop-in that sorts after Omarchy's own defaults. Use tee -a, because a bare tee truncates the file and would discard any parameter you added earlier from another record:
sudo tee -a /etc/limine-entry-tool.d/zz-local.conf >/dev/null <<'EOF'
# `+=` appends; a bare `=` would wipe Omarchy's defaults.
KERNEL_CMDLINE[default]+=" amdgpu.lockup_timeout=10000"
EOF
sudo limine-mkinitcpio && sudo limine-update && sudo reboot
Useful values, per amdgpu's own module documentation:
amdgpu.lockup_timeout=10000: watchdog in ms (default 2000). Format is a single value orGFX,Compute,SDMA,Video. Use only if a legitimately long compute/shader job is being killed.amdgpu.noretry=0: re-enable XNACK retry faults (0 = retry enabled, 1 = retry disabled, -1 = auto). Worth a try on GFX9/Vega where a VM fault escalates into a ring hang.amdgpu.gpu_recovery=1: **only useful on legacy ASICs.** The-1auto default already enables recovery on everything except SI, CIK, Carrizo, Stoney and Cyan Skillfish, so on RDNA2/RDNA3 this parameter changes nothing. If your reset is failing witherror -110, that is a firmware/hardware problem and this will not fix it.
On GRUB systems put the same parameters in GRUB_CMDLINE_LINUX_DEFAULT in /etc/default/grub and run sudo grub-mkconfig -o /boot/grub/grub.cfg.
Verify
Run the offending workload for 30+ minutes. journalctl -k -b | grep -i 'ring .* timeout' stays empty, or at worst shows but soft recovered with no GPU reset. cat /sys/class/drm/card1/device/gpu_recovery 2>/dev/null and cat /sys/module/amdgpu/parameters/gpu_recovery reflect the value you set.
Sources
- https://github.com/torvalds/linux/blob/master/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
- https://github.com/torvalds/linux/blob/master/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
- https://docs.mesa3d.org/envvars.html
- https://bbs.archlinux.org/viewtopic.php?id=301378
- https://bbs.archlinux.org/viewtopic.php?id=284033
- https://wiki.archlinux.org/title/AMDGPU