Been tinkering with NixOS specializations trying to optimize my portable gaming setup on the framework 13 Ryzen AI 7 350. Sharing the results here just for fun.
The System
Mainboard: Ryzen AI 7 350
RAM: 64GB DDR5 5600 (2x32GB, Crucial)
Dedicated VRAM: 16GB (via BIOS setting)
Kernel: 6.17.7-cachyos, compiled with x86_64-v3 optimizations.
Proton: proton-cachyos-10.0-20251023, with x86_64-v4 optimizations.
Filesystem: Btrfs
SSD: 2TB SK Hynix Platinum P41
Screen: The built-in 2.8k 120Hz screen.
Render Resolution: 1920x1280 (via in-game video settings)
Power Limit: 30w*
Fan speed: 100% (via fw-fanctrl)
* Any higher and RAM overheats, crippling the framerates by over 20%. Framework, please get better at keeping your RAM cool!
The Desktops
KDE
Sort of a “control” case, because this is a hella popular DE that many many people use for gaming. I’m on KDE Plasma 6.5.1.
Cosmic
The new DE from the pop_os! folks, all in rust with cool auto-tiling. Using the latest beta 4 release.
Gamescope
Yes, a raw gamescope session, not nested in any other DE. Note that this is the only DE (if you can call it that) where I had to launch the game in xWayland, because the game simply would not launch when using gamescope’s --expose-wayland argument. Version 3.16.17.
Metheds
On a clean reboot, launch cyberpunk from Lutris with native wayland and ntsync enabled via environment variable. Fullscreen the game, and immediately run the built-in benchmark. Repeat 3x for each desktop, and record the highest and lowest runs for each.
The benchmark was run on medium settings with all post processing effects disabled (vignetting, motion blur, etc.). FSR 3.0 scaling was set to balanced, and FSR 3.1 frame generation was enabled.
Results
| KDE | Cosmic | Gamescope | ||
|---|---|---|---|---|
| Min FPS | 59.08 - 61.38 | 60.83 - 60.90 | 51.55 - 51.49 | |
| Avg FPS | 68.00 - 70.21 | 69.29 - 69.22 | 57.27 - 57.01 | |
| Max FPS | 78.48 - 80.58 | 78.93 - 78.77 | 63.75 - 63.51 |
Conclusion
Surprisingly, the gamescope session did pretty poorly. I suspect either this game or this proton version really dislikes xWayland. Maybe gamescope really is just slow, but I’m inclined to give valve the benefit of the doubt and assume xWayland might be the culprit in this particular title.
KDE and Cosmic were neck-and-neck in terms of averages, with the two DEs performing within margin of error of each other. KDE’s results were slightly more varied than Cosmics, with good runs shining brighter but bad runs sinking farther as well. I actually re-ran KDE’s a few more times after finishing my testing to see if the variations between runs would settle, but it stayed pretty volatile across many runs. Not sure why, maybe some background shenanigans.
I gotta give cosmic a narrow win over KDE for incredible consistency, but you wouldn’t be missing much with either choice.