Re:Graphics card generation is over and it was mostly trash
by Aardwolf on Monday September 11, 2023 @09:47AM
(#37467205)
> This is also why it requires NVIDIA’s bespoke Tensor cores to function.
Any idea why this needs bespoke cores? GPU's were already used for AI and crypto mining before this, without bespoke cores for 1 specific application
> The “Deep Learning SuperSampling” is right in the title
With that name, it could have been based off of a single pre-training on many games, to work on games in general
Re:Graphics card generation is over and it was mostly trash
by gjsman-1000 on Monday September 11, 2023 @09:50AM
(#37467238)
As far as I know, Tensor cores are a little like Raytracing cores. Precision, unlike for mining or science, is not valued and they are allowed to do sloppy math within a tolerance in exchange for much, much greater speed.
> With that name, it could have been based off of a single pre-training on many games, to work on games in general
Well, I’m not NVIDIA, so I don’t know if you could. I just know AMD tried and the results were almost universally considered inferior.
Because DLSS is a AI-based solution with customized settings and inputs for each game. (The “Deep Learning SuperSampling” is right in the title.) This is also why it requires NVIDIA’s bespoke Tensor cores to function. DLSS 2.0 removed most per-game optimization but there’s still some support work required.
AMD’s FSR is basically a traditional upscaling solution, just with low overhead. FSR runs fine on NVIDIA and even the Nintendo Switch - but almost all reviewers agree DLSS is superior in most games.
Except, perhaps, Cyberpunk 2077. But that’s not a fair metric to measure DLSS by because everything about that game has been a technical disaster.