[Nvidia] DLSS boosts Port Royal Benchmark performance by up to 50%

https://www.nvidia.com/en-us/geforce...yal-benchmark/

I like to hear what OCN has to say about this. I personally think this benchmark is a sham. I might change my mind depending on how BF5 and Anthem run with DLSS.

Predictive AI enhancement for a static benchmark is meaningless. Couldn't care less about this. Until its in a game with ray tracing its just a glorified AA solution.

DLSS is natively rendering it at 1080P then uspcaling to 1440P, of course i'd get a large FPS boost.
the question is whether the quality is up to par.

Isn't it supposed to be just a glorified AA solution? In fact, that is what it would be best at IMO, give me an image rendered at 4K and use the neural net to do good AA without needing to super sample. 😛

Doesn't it actually look better with DLSS from that video?

isn't it the opposite of supersampling? in supersampling the screen is rendered at a higher resolution then downscaled with a loose scaler.

while DLSS on the other hand uses a lower resolution then upscaled with a smart scaler.



probably because the render benchmark had predetermined frames, where the trained AI is already aware of what to adjust to make it better.

ULmark and their DLSS demos are load of rubbish. The TAA I've seen them implement is one of the worst ever and who the hell uses TAA in first place when there is MSAA, SMAA, TSSAA8TX, etc.
Compare 1440p vs 1440pDLSS and 4k vs 4kDLSS, not these fake4k upscaling where it's comparing 4k vs 1440pDLSS, of course 1440pDLSS is faster than rendering native 4k, duh!

Nvidia probably still waits to enable the DLSS2x which is the native DLSS mode without upscaling. They likely know every well it's slower or not really beneficial over traditional AA methods.

Exactly. This is what I was trying to say in my post.

DLSS doesn't have to use a lower resolution, Nvidia could simply train a neural network to do great AA. When we first heard about it this is what it sounded like. Lower resolution is faster but I bet a neural network could offer great AA quality faster than 4xSSAA working with the full resolution.

Yes, but there is only one AI and a lot of frames, so it still has to work well on all the frames with only one set of weights. Nvidia is probably not using a separate neuralnet for each frame.

The variety of frames is still really low compared to a game due to the fixed nature of the benchmark, but it is more a best case for DLSS rather than completely cheating.