AMD FSR upscaling

DavidGraham · Wednesday at 9:34 PM

Lurkmass said:
Well on their "custom driver" (RADV), they haven't implemented/exposed FP8 support for GFX12 so they use FP8 'emulation' (FP8 <-> FP16 conversion) like I mention above and it gives us an apples to apples comparison as well feature use wise ...

So, on Windows, FSR4 on RDNA4 should be even faster than this. It shouldn't be surprising if 7900XTX turns out to be a 100% slower.

LordEC911 · Thursday at 3:14 AM

The screenshots also show vastly different CPUs... a 7800x3d w/ 7900xtx vs a 3950x w/ 9700XT(GFX1201).

MfA · Thursday at 11:40 AM

Lurkmass said:
Yes, they convert FP8 values to FP16 values and back since there's no official Vulkan extension which exposes native FP8 tensor formats ...

A quick custom extension and hacking it into RADV isn't much more work, but I guess that's not his job.

PS. Not to say it's easy, but given what he did in two days ...

Cyan · Thursday at 8:37 PM

FSR4 is available in Linux thanks to a person who reverse engineered it. The guy who implemented this made it work not only on RDNA 4 GPUs but on the previous generation.