Discussion Nvidia Blackwell in Q1-2025

[adroc_thurston]

I'm not sure using the same base resolution and measure FPS count is the best way to compare upscalers.

There should be some effort to reach ISO-quality regardless of base resolution and only then should they measure average FPS and/or frametimes. I'm not saying this is the case, but DLSS4.5 with base 1080p for a 4K output could show identical IQ to DLSS4 with base 1440p for the same 4K output, and performance on the former could be better.

Nonsense, they're selling this as an IQ improvement for the existing DLSS install base.
If you have to lower down your preset to maintain perf (and thus sacrifice the IQ) then the update is a meme.

 

[basix]

When you read the docs: The new presets L and M are indeed much heavier

DLSS/doc/DLSS_Programming_Guide_Release.pdf at main · NVIDIA/DLSS

NVIDIA DLSS is a new and improved deep learning neural network that boosts frame rates and generates beautiful, sharp images for your games - NVIDIA/DLSS

github.com

And from the performance numbers including memory usage they seem to have switched to and FP8 based DNN (at least in some parts). Turing and Ampere performance falls apart because sticking to FP16.
Model M seems to trade execution speed with a little bit (tiny bit) of memory overhead. Let's see what the quality differences will be.

And hey: Creating a selling point for Rubin with much higher Tensor throughput sounds like business strategy

 

[SolidQ]

Even RTX 5060 drop fps

 

[CakeMonster]

I'm worried about what I read about sharpening, I hope that's optional or there are presets for us who don't like any visible trace of that type of effect.

 

[jpiniero]

I'm worried about what I read about sharpening, I hope that's optional or there are presets for us who don't like any visible trace of that type of effect.

The older presets are going to stick around.

 

[MrMPFR]

Even RTX 5060 drop fps

2-3% overhead. Nice try NVIDIA. Not everyone has a 5090.

I'm worried about what I read about sharpening, I hope that's optional or there are presets for us who don't like any visible trace of that type of effect.

The new models are tuned for performance and ultra performance only. Use older models for everything else.
But hey it can make 720P UP usable:

 

[MrMPFR]

IIRC CPX extrapolation was 8x raw NVFP4 dense vs 50 series. IF DLSS (currently FP8) works with NVFP4 then that's 16X matmul at iso-core + clocks.

Nvm. Kopite7kimi seems to have killed that.
CPX = 6090 die with all gfx stripped out except ROPS for AI slop generation.