Low GPU Usage % in games using Compute functions (PhysX)

Using a 3060ti, I’ve noticed performance being distressingly low on some games in Proton when compute functions like PhysX are enabled. I can specifically reproduce this on Batman: Arkham Asylum and Arkham City, where the former (despite being the older title, 2008 vs. 2011) performs significantly worse in relatively simple scenarios if some PhysX objects are in the scene - but this problem affects the latter as well.

(Arkham Asylum - Challenge Map 1; has no formal benchmark, but 7%(!) usage)

(Arkham City - Benchmark, Scene 2)

(These are results after several runs of both scenarios, negating performance penalties due to shader compilation. Yes, Arkham Asylum dips to 7% usage and 18 FPS in gameplay and maintains that.)

Where one would expect these to be because of high GPU usage (unlikely, considering the age of said games relative to the card), in actuality when PhysX is enabled, the GPU is using less resources - almost like something is intentionally throttling the performance of the GPU, or it isn’t detecting the game process properly to recognize that it should be using max power. When PhysX is disabled for these games, they will indeed run as fast as possible.

Both of these games use an older accelerated PhysX implementation (not sure of the specifics, but the configuration windows for both these games specifies they only work on NVIDIA GTX 400/700-series). This was also observed with nvidia-smi --lock-gpu-clocks=2160,2160 (my GPU’s maximum), effectively forcing the 3D clocks to its reported max.

Using Arch Linux, Proton Experimental, Prop. Driver 535.113.01, RTX 3060 ti, Ryzen 5 5600X, 16GB RAM. If there’s any other clarifications needed, I’ll be happy to offer.