sc94597 said:
Looking into it, this occurred especially in the new implementation with DLSS 4. The old RR model seemed to have about a 6% performance penalty on an RTX 3080 in CP2077 and an 11% penalty on a 3060. About 12% for a 3050. So yeah, more significant than I remembered, even with the old model. S2 has about 60% of the tensor cores of the 3050 desktop, so probably not doable, unless they could do a lightweight/distilled model. |
Yeah, maybe they can do some lighter model, like with DLSS (still waiting to see what DF digs out from devs about that one).







