The Witcher 3 Remastered: whole-system freeze / amdgpu "device lost from bus" on RX 6800 XT when starting a new game

vkd3dopen appid 292030needs feedback

HansKristian-Work/vkd3d-proton#3347 · opened 2026-10-03 by GodlikeIDGAF · updated 2026-10-26 · 3 comments · github · game page · search this game

GGodlikeIDGAF 2026-10-03 github

The Witcher 3 Remastered (292030): whole-system freeze / amdgpu "device lost from bus" on RX 6800 XT when starting a new game

Summary

Since the Remastered update (Steam build 25646871, installed 2026-10-02), starting a New Game freezes the entire machine (screen signal lost, PC still powered). The game worked before the update, in early September 2026, on the same machine and software stack. The main menu is fine. The freeze happens at the first 3D scene (the Yennefer/Geralt intro cinematic), reproducibly: about 12 out of 12 attempts.

The kernel log sometimes survives the freeze and shows the GPU dropping off the PCIe bus, with vkd3d_queue of witcher3.exe as the offending process.

I have ruled out hardware and system causes as far as I can (details below). I think this is the same family as the other reports of crashes after the Remastered / v5.00 update on Proton (ValveSoftware/Proton#10213, nanomatters/proton-cachyos#71), but the symptom here is more severe (GPU lost, not just a game crash).

System

GPU: AMD Radeon RX 6800 XT (Navi 21, SIENNA_CICHLID, PCI 0b:00.0), Mesa RADV
CPU: Ryzen 7 3700X, board: ASUS TUF GAMING X570-PLUS (WI-FI), BIOS 5021 (2024-09-29)
RAM: 2x32 GB Corsair CMK64GX4M2E3200C16 at 2133 MT/s (no XMP/DOCP)
PSU: Antec HCG850 (850 W, 2021)
OS: CachyOS, KDE Plasma 6.7.5 on Wayland
Kernel: 7.2.8-1 and 7.2.8-2-cachyos; also reproduced on 6.18.52-1-cachyos-lts
Mesa / vulkan-radeon: 26.2.4 (26.2.3 on the very first crashes); linux-firmware 20260916
Steam 1.0.0.87-3; Proton: GE-Proton11-7 and Proton Experimental (experimental-11.0-20261001), same result
Game files on an NTFS disk (ntfs-3g / fuseblk); Proton prefix on btrfs (STEAM_COMPAT_DATA_PATH)
Launch options for the tests: STEAM_COMPAT_DATA_PATH="..." %command% --launcher-skip -disablemods (no OptiScaler, no NVAPI, no VKD3D_CONFIG=dxr, no gamemode, no MangoHud, Steam overlay and shader pre-caching disabled)
Game settings at the time: ray tracing off, no upscaler, no frame generation (EnableRT=false in dx12user.settings), 3440x1440. Also reproduced at the Low preset, 1080p, HairWorks off.

Kernel log when it is captured (`journalctl -f -k` running during the test)

amdgpu 0000:0b:00.0: Dumping IP State
amdgpu 0000:0b:00.0: device lost from bus!
amdgpu 0000:0b:00.0: SMU: bus error for message: DisallowGfxOff(41) response:0xFFFFFFFF
amdgpu 0000:0b:00.0: Failed to disable gfxoff!
amdgpu 0000:0b:00.0: [drm] AMDGPU device coredump file has been created
amdgpu 0000:0b:00.0: ring gfx_0.0.0 timeout, signaled seq=153954, emitted seq=153957
amdgpu 0000:0b:00.0:  Process MainThread pid 9483 thread vkd3d_queue pid 9880
amdgpu 0000:0b:00.0: Starting gfx_0.0.0 ring reset
amdgpu 0000:0b:00.0: Ring gfx_0.0.0 reset failed
amdgpu 0000:0b:00.0: GPU reset begin!. Source:  1
amdgpu 0000:0b:00.0: GPU reset end with ret = -19
amdgpu 0000:0b:00.0: GPU Recovery Failed: -19

The same sequence on 6.18.52-lts:

amdgpu: SMU: response:0xFFFFFFFF for index:18 param:0x00000005 message:TransferTableSmu2Dram?
amdgpu: Failed to export SMU metrics table!
amdgpu: device lost from bus!

In other runs the whole system froze with nothing written to the journal.

PCIe link looks healthy when idle: LnkSta: Speed 16GT/s, Width x16, no AER errors.

Power / temperature trace at the moment of the freeze (1 Hz, `sensors`)

Last samples before the GPU is lost or the log stops (junction temp / PPT):

Run 1: 54 C / 23 W, then 58 C / 107 W, then sensors N/A (GPU gone).
Run 2 (Proton Experimental): 56 C / 32 W, then 57 C / 90 W, 59 C / 115 W, 72 C / 224 W, 75 C / 261 W, log ends.
Run 3 (Low preset, 1080p): about 65 W and 47 C for roughly 90 s of intro cinematic, then the log ends.

So the freeze happens at low and high power alike, with low temperatures (never above 77 C junction, limit 110 C).

What I ruled out

Kernel version: same on 7.2.8-1, 7.2.8-2 and 6.18.52-lts.
amdgpu.gfxoff=0, forced power_dpm_force_performance_level=high (no clock state transitions), clocks capped below the top DPM level: no change.
Proton version: GE-Proton11-7 and Experimental behave the same.
Steam overlay, shader pre-caching, MangoHud, gamemode, OptiScaler, NVAPI, DXR: all removed for the tests.
Cleared shadercache/292030 and vkd3d-proton.cache, ran "Verify integrity of game files" (Steam redownloaded a few files).
Hardware under load: FurMark (Vulkan, 1080p) ran for 10 minutes at 264 W / 84 C junction with no error; the load stepped from 23 W to 263 W in about 2 s, steeper than the game's. vkmark also passes. GPU, PSU, cables and PCIe slot look fine.
RAM at stock speed, no thermal issue (idle junction about 55 C).

Expected / actual

Expected: the New Game intro loads and plays, as it did before the update.
Actual: the whole system hangs a few seconds to about 1.5 minutes after the first 3D scene starts; reduced settings only delay it.

Questions

Is this a known regression of the Remastered build on RDNA2 with vkd3d-proton / RADV?
Is there a way to get a useful trace (e.g. VKD3D_DEBUG, RADV_DEBUG=hang) given that the machine hard-freezes? I can run anything you suggest with the kernel log and a 1 Hz sensors log running.

Related reports: ValveSoftware/Proton#10213 (menu and cutscene crashes after the Remastered update, RX 9070 XT), nanomatters/proton-cachyos#71 (most Proton builds crash at startup after the update).

GGodlikeIDGAF 2026-10-05 github

Update: still reproducible on a clean install, ext4, kernel 7.2.9

To rule out my storage setup, I reformatted the game disk from NTFS to ext4, reinstalled the game from scratch (fresh files, fresh Proton prefix), and updated the kernel. Same result:

Kernel 7.2.9-1-cachyos, Mesa 26.2.4, GE-Proton11-7, game on ext4, new game started at the Low preset
About 10 seconds after the first 3D scene starts, the GPU drops off the bus:

amdgpu 0000:0b:00.0: device lost from bus!
amdgpu 0000:0b:00.0: SMU: bus error for message: TransferTableSmu2Dram(18) response:0xFFFFFFFF
amdgpu 0000:0b:00.0: Failed to export SMU metrics table!
amdgpu 0000:0b:00.0: ring gfx_0.0.0 timeout, signaled seq=66723, emitted seq=66726
amdgpu 0000:0b:00.0: Process MainThread pid 8656 thread vkd3d_queue pid 9126

Just before the loss: about 110 W / 62 C junction (1 Hz sensors log), so this is not a power or thermal problem.

Previously ruled out: kernels 7.2.8-1, 7.2.8-2 and 6.18.52-lts, Proton Experimental (experimental-11.0-20261001) and GE-Proton11-7/11-6, amdgpu.gfxoff=0, forced high DPM level, cleared shader caches, and a 10-minute FurMark Vulkan stress test at 264 W (stable). vkd3d-proton builds bundled: GE-Proton11-7 is 5604-gaf89350cc, Experimental is 5641-g44cf7c20 (3.1.0).

I can run any debug build or trace you want (with the kernel log and a 1 Hz sensors log running), but the machine hard-freezes on failure, so output may be lost.

Kkarlschriek 2026-10-20 github

Same symptom on a second Navi 21 card.

Game audio kept playing after the GPU dropped off the bus, so userspace was still running.

Hardware / software


GPU	MSI RX 6900 XT, Navi 21, `1002:73bf` rev `0xc0`, subsystem `1462:3952`
Distro	Ubuntu 26.04.1 LTS
Kernel	7.0.0-38-generic
Mesa	26.0.8 (RADV)
linux-firmware	20260319.git217ca6e4
SMU	`driver if 0x40, fw if 0x41, version 0x003a5b00 (58.91.0)`
`ppfeaturemask`	`0xfff7bfff` (kernel 7.0 default)
PCIe link	16.0 GT/s x16, card directly in the slot, no riser
Game	The Witcher 3: Wild Hunt Remastered, app 292030, build 25773555

When it started

The game ran fine on this exact build for eight days before the first crash.

11 Oct: game updated to build 25773555 (Remastered). Played without problems on Ubuntu 22.04, kernel 6.8, Mesa 25.0.5.
19 Oct: upgraded 22.04 to 26.04. Kernel 6.8 to 7.0, Mesa 25.0.5 to 26.0.8, linux-firmware to 20260319.
20 Oct: first crash, and every attempt since.

So on my system the Remastered update alone was not sufficient to trigger it. Something in the newer stack is the other half.

Telemetry across the failure

GPU sensors logged once a second, flushed per line:

time         busy  junction  edge  power  fan   sclk
13:28:33.90   85%   59 C     52 C  142 W  1243  579 MHz   <- last healthy sample
13:28:34.99   --    0        0     0      --    0         <- all sensors gone, VRAM frozen

59 C and 142 W at the moment it died, against a 110 C limit and a 300 W budget. Not thermal, not a power-limit event.

Kernel log

amdgpu 0000:0b:00.0: device lost from bus!
amdgpu 0000:0b:00.0: SMU: bus error for message: AllowGfxOff(40) response:0xFFFFFFFF
amdgpu 0000:0b:00.0: Failed to enable gfxoff!
amdgpu 0000:0b:00.0: device lost from bus!
amdgpu 0000:0b:00.0: SMU: bus error for message: TransferTableSmu2Dram(18) response:0xFFFFFFFF
amdgpu 0000:0b:00.0: Failed to export SMU metrics table!
[... repeats, 695 lines total ...]

Nothing precedes it. No ring timeout, no GPU reset begin, no MES failure, no PCIe AER errors. The lines immediately before are unrelated dbus audit messages. No GPU reset is attempted at all.

Already ruled out here

GFXOFF. Worth knowing if you are still testing amdgpu.gfxoff=0: on kernel 7.0 the default ppfeaturemask is 0xfff7bfff, which already has the GFXOFF bit clear. The card dropped with GFXOFF off, so that parameter is a no-op on this kernel.
Ray tracing. EnableRT=false, RTGIPreset=false, EnableRtRadiance=false, RTAOEnabled=false in dx12user.settings, all set before the crashes began.
Runtime PM. Never engages here. runtime_suspended_time stays 0 with three displays attached.
PCIe. Clean Gen4 x16, zero AER errors, no riser.
Reseating the card. No change.
Thermal and power. 59 C junction, 142 W at failure.

Rrunar-work 2026-10-26 github

Do you still have this issue? I haven't been able to reproduce this so far with RX 6800, 7.2.11-arch1-1, Mesa 26.2.5, 26.0.8 and main (ecd75369), Proton Experimental, various graphics presets (low, ultra, ultra RT).