RX 9070 XT: games running across suspend lose VRAM to GTT on resume, performance collapses until restart

[jedhamzawi]

Your system information

• Steam client version: 1789086785 (steamdeck_publicbeta)
• SteamOS version: 3.9.0 (build 20260828.100), kernel 7.2.0-valve1. Also reproduced on 3.8.26 (20260827.1), kernel 6.18.46-valve1
• Opted into Steam client beta?: Yes
• Opted into SteamOS beta?: Yes (Preview)
• Have you checked for updates in Settings > System?: Yes

Hardware: desktop, ASRock B850I Lightning WiFi (BIOS 4.43), Ryzen 7 7800X3D, XFX RX 9070 XT (Navi 48, 1002:7550), HDMI to a 4K HDR 120Hz TV. Raphael iGPU is also bound to amdgpu.
Graphics stack: vulkan-radeon 26.2.0_devel.222131.steamos_26.05.16, mesa 26.1.2, linux-firmware-neptune 20260827.1, gamescope 3.16.26.

Please describe your issue in as much detail as possible:

Expected: a game left running when the system goes to sleep performs the same after wake as before.

Actual: after wake the game is severely degraded until it is restarted. A game launched after wake runs normally. Reproduced with:

• Onimusha: Way of the Sword
• The Witcher 3
• Counter-Strike 2 (native Linux/Vulkan)

I have two OS versions I've tested that fail differently:

3.9.0: frame rate collapses (Onimusha drops from 60fps to about 12fps). The game's VRAM is evicted to GTT at suspend and never comes back:

• Before suspend, Onimusha's fdinfo shows drm-resident-vram at about 7.4 GiB and mem_info_vram_used at 8.6 GiB.
• After wake, Onimusha shows drm-resident-vram at about 190 MiB and drm-resident-gtt at about 7.9 GiB. mem_info_vram_used is about 1.0 GiB and mem_info_gtt_used about 8.3 GiB.
• CS2 after wake: mem_info_vram_used is 1.8 GiB and mem_info_gtt_used 8.5 GiB, almost all of it CS2's. CS2 relaunched after the same wake holds a solid 60fps with mem_info_vram_used at 9.8 GiB and mem_info_gtt_used at 0.5 GiB.
• It stays that way for as long as the game runs. The GPU reports 100% busy, but power drops from about 230-266W to about 140-170W.
• No kernel errors are logged.

3.8.26: VRAM does return, but the game hitches badly on camera movement. Both measurements below were taken while the game was running after wake:

• The game's gfx engine time is about 2.8x its pre-suspend value, and power falls to about 120W at 100% busy.
• amdgpu fires about 85k IRQs per second. kworkers on the sdma1, gfx_0.0.0 and comp_1.1.0 workqueues stay busy.
• mem_info_vram_used / mem_info_gtt_used change between 20Hz samples about half the time (±30 MiB).

Both kernels log amdgpu 0000:03:00.0: MODE1 reset at suspend entry.

On 3.9.0 I ruled out each of the following with a separate suspend cycle; the result was identical every time:

• my own GPU tuning, both the undervolt/power cap/fan curve and OverDrive disabled at module load via stock ppfeaturemask
• dmemcg-booster, with both units masked and the dmem controller not enabled
• Steam background game recording
• the iGPU: the game's memory stays behind the dGPU's GTT, and the iGPU holds about 14 MiB

Sleep/wake worked on this system at some point, but I don't know which version last worked.

Steps for reproducing this issue:

1. On an RX 9070 XT, launch Counter-Strike 2 (offline practice map, 4K), Onimusha: Way of the Sword, or The WItcher 3 and play briefly.
2. Put the system to sleep with the game still running, then wake it.
3. Return to the game: on 3.9.0 performance collapses and /sys/class/drm/card0/device/mem_info_vram_used stays around 1-2 GiB with the game's memory in GTT; on 3.8.26 it hitches on camera movement. Restarting the game restores normal performance.

Other Notes

I haven't found others with this specific issue, so it could be something specific to my setup. That being said, I have pretty well-supported hardware and am running pretty much stock SteamOS besides a small undervolt and power limit settings (bug manifested with that disabled though, so that's not the problem). This bug in particular is a huge bummer for me because sleep/resume is the most important feature for me in SteamOS. I would appreciate any guidance/advice/possible fixes!

[matte-schwartz]

does it happen if you play the games on desktop mode and suspend there? or only in game mode.

[jedhamzawi]

The actual process of achieving sleep in desktop mode is a bit weird for my system. I have the big picture mode interface enabled, and when I go to start sleep in the UI, the Steam logo appears to show it's sleeping, the game goes black, but the system does not sleep. I have to press the power button to actually force start sleep mode. This is not the case in game mode.

In any case, I have the same wake behavior in desktop mode. VRAM usage before sleep is at about 8GB and is a bit over 1GB after waking. FPS is at 10-15. So it doesn't seem to be gamescope/game mode specific. This is in onimusha.

[matte-schwartz]

I can repro the issue in Onimusha on Steam Machine, just downloaded the game to check and it went from locked 60fps in the starting cutscene to 14fps after suspend/resume. Will look into it.