[BUG] Claude Desktop Linux triggers amdgpu GFXHUB page faults and resets GNOME Wayland session

What's wrong?

The official Claude Desktop Linux app intermittently submits a GPU workload that triggers an amdgpu GFXHUB page fault on an AMD Raphael iGPU. The kernel identifies claude-desktop as the process owning the faulting GPU context. This is followed by a graphics ring timeout and, in three of four observed occurrences, a full GPU reset that terminates the GNOME Wayland session. One occurrence was followed by an abrupt reboot.

This happened four times within about five hours on July 29, 2026, across three different Claude Desktop GPU-process PIDs. I searched the existing issues and did not find an equivalent Linux/AMDGPU report.

The fault may ultimately be in Electron/Chromium, Mesa, or amdgpu, but Claude Desktop consistently owns the initial faulting context.

What should happen?

Claude Desktop hardware acceleration should not wedge or reset the system's primary GPU. If the GPU context fails, the app should recover without terminating the desktop session.

Environment

• Claude Desktop: 1.24012.9 (amd64, official Anthropic .deb)
• Installed: July 27, 2026 (no known previous working version)
• OS: Ubuntu 26.04 LTS
• Session: GNOME 50.1 / Mutter 50.1, Wayland
• Kernel: 7.0.0-28-generic (Ubuntu 7.0.12)
• Mesa: 26.0.3-1ubuntu1
• Primary/display GPU: AMD Raphael iGPU [1002:164e], amdgpu
• Secondary GPU: NVIDIA GeForce RTX 5070 Laptop [10de:2d18]
• Internal panel was driven by AMD during all four incidents

Steps to reproduce

The trigger is intermittent rather than tied to one known UI action:

1. Start Claude Desktop normally on a GNOME Wayland session where the AMD iGPU drives the internal panel.
2. Use the app for an extended session with hardware acceleration enabled.
3. The claude-desktop GPU context eventually generates a GFXHUB page fault.
4. amdgpu reports a ring timeout; the GPU context or the whole GPU is reset.

Observed occurrences:

| Time (local) | Initial fault owner | Result |
|---|---|---|
| 2026-07-29 14:01:40 | claude-desktop PID 7851 | Ring reset succeeded; session survived |
| 2026-07-29 15:50:38 | claude-desktop PID 7851 | Full GPU reset, VRAM lost, Claude GPU process exited, GNOME session terminated |
| 2026-07-29 18:35:36 | claude-desktop PID 8083 | Full GPU reset, GNOME Shell aborted, session terminated |
| 2026-07-29 19:17:27 | claude-desktop PID 20420 | Full GPU reset; journal ended shortly afterward and the next boot began |

I checked the retained system journal from July 23 through July 30. These are the only AMD GFXHUB page faults/ring resets in that period, and every initial page fault is attributed to claude-desktop.

Sanitized kernel log

Representative full-reset occurrence:

Jul 29 15:50:38 kernel: amdgpu 0000:05:00.0: [gfxhub] page fault (src_id:0 ring:24 vmid:5 pasid:61)
Jul 29 15:50:38 kernel: amdgpu 0000:05:00.0: Process claude-desktop pid 7851 thread claude-des:cs0 pid 7908
Jul 29 15:50:38 kernel: amdgpu 0000:05:00.0: GCVM_L2_PROTECTION_FAULT_STATUS:0x00501430
Jul 29 15:50:38 kernel: amdgpu 0000:05:00.0: Faulty UTCL2 client ID: SQC (data) (0xa)
Jul 29 15:50:38 kernel: amdgpu 0000:05:00.0: PERMISSION_FAULTS: 0x3
Jul 29 15:50:40 kernel: amdgpu 0000:05:00.0: ring gfx_0.1.0 timeout, signaled seq=3297583, emitted seq=3297585
Jul 29 15:50:40 kernel: amdgpu 0000:05:00.0: Process gnome-shell pid 3423 thread gnome-shel:cs0 pid 3562
Jul 29 15:50:40 kernel: amdgpu 0000:05:00.0: Ring gfx_0.1.0 reset failed
Jul 29 15:50:40 kernel: amdgpu 0000:05:00.0: GPU reset begin!. Source: 1
Jul 29 15:50:40 kernel: amdgpu 0000:05:00.0: MODE2 reset
Jul 29 15:50:40 kernel: amdgpu 0000:05:00.0: GPU reset succeeded, trying to resume
Jul 29 15:50:40 kernel: amdgpu 0000:05:00.0: VRAM is lost due to GPU reset!
Jul 29 15:50:41 kernel: amdgpu 0000:05:00.0: GPU reset(2) succeeded!
Jul 29 15:50:41 kernel: amdgpu 0000:05:00.0: [drm] device wedged, but recovered through reset
Jul 29 15:50:41 kernel: amdgpu 0000:05:00.0: [drm] *ERROR* Failed to initialize parser -125!

Claude/Electron immediately afterward:

SharedContextState context lost via EXT_robustness. Reset status = GL_INNOCENT_CONTEXT_RESET_KHR
Restarting GPU process due to unrecoverable error. Context was lost.
GPU process exited unexpectedly: exit_code=8704
Error reading events from display: Broken pipe

At 18:35, the same sequence ended with:

org.gnome.Shell@ubuntu.service: Main process exited, code=dumped, status=6/ABRT

Additional information

• Reinstalling the same 1.24012.9 package after the second and third incidents did not prevent recurrence.
• System memory was not exhausted during the representative event.
• I have not yet tested --disable-gpu, because that would avoid rather than isolate the failing accelerated path.
• Raw Claude logs are not attached because they contain account metadata; the excerpts above are sanitized.