Darc Requiem said:
From my understanding, the Linux driver for AMD has an FP16 to FP8 translation layer that is not in the Windows drivers. So technically FSR4 is using two translation layers on Linux for RDNA3. The FP16 To FP8 translation layer and the Proton translation layer. |
Your understanding is correct. They are just translating FP8 to FP16... And this is only possible because RDNA3 supports Wave Matrix Multiply Accumulate on the FP16 data path.
Keep in mind that the Radeon 7900XTX has 61 Teraflops of FP16 available, which is why it can brute force this...
I would imagine a lower-end GPU with far less FP16 capacity would see a reduced return on investment.
www.youtube.com/@Pemalite
