| JEMC said: I see. The performance loss is huge, yes, but since FSR4 is roughly one month old, I'd say that this project is still in the early stages and that there's room for improvement, even if they'll never reach parity. |
Mind that there is no native FSR4 on Linux, so you need to:
a) emulate all of AMD's new proprietary WMMA instructions in the first place, which is a bit slow and looks bad at the moment even on RDNA4;
b) then use a hack to compile for yourself a custom version that supports RDNA3, since that can't be done by default with the software renderer being used (Mesa 3D) because it doesn't support these sorts of instructions yet.
Still, I'd call bull on that number, that would be slower than software ray tracing and that requires emulating much more complex BVH math instead. Unless that is referring to the total time the upscaler is taking (say, 14ms vs. 2ms) which is a more reasonable overhead for this sort of thing.
