50 series isn't a new architecture over the 40 series though, it's just revised of current gen.
That's not accurate though. With CUDA cores, sure, nothing much changed, raster is same just more cores and higher clocks. But RT and AI processing got a very nice boost:
1. With Tensors we've got native FP4 and FP6 support, so inference in those native formats is considerably faster than on 4k series (FP8 native only). It also got native de-quantization which allows on the fly processing of such formats without using expensive FP16. That plus faster vRAM and gigantic cache on 5090 and it can be much faster in AI processing than 4090 for example, way faster than clocks, power and number of tensor cores would suggest. IF the model is optimised for it. And this is the funny bit - DLSS5 does NOT seem to be optimised for it, it's FP8, so 4k series and older. Very curious! It could've been done because of the quality drop below 8bits, though, hard to say.
2. With RT we got 2x higher ray-triangle intersection throughput (BVH etc.). To that "Mega Geometry", which is native hardware acceleration for clustered geometry (4k and older can't do that, they go per individual triangles). It's a bit similar to nanite in UE5, but actually speeds things up in practice, instead of just in theory.
They also added linear swept spheres that use hardware acceleration for raytracing of curved primitives like hair etc. without expensive approximations like 4k and older had to do. And most importantly in 5k series they added two-way low latency execution loop between CUDA and RT cores, which allowed them to implement things like SER, micro-mesh & displaced micro-geometry feedback (helps with RT of dense geometry things like fur and leaves etc.) - and all that stuff runs in the internal cache, never hitting vRAM, which makes it WAY faster. And as last thing, they added in software APIs (like DXR 1.2 / Shader Model 6.9) native HitObject pipeline hooks. This allows shader code running on CUDA cores to hand off incomplete rays mid-execution back to the RT Cores, wait for a quick hardware check, and pick up the result in the exact same execution pipeline without restarting the ray-generation pass. Again, WAY faster to execute than ever before.
So, generally, very nice improvements that... are almost invisible in games as hardly any game implemented those, as engines do not expose those and modern devs can't write proper code for nothing, it seems (or even tweak knobs in the engine!). Things like that are expensive to implement as you actually need to hire senior devs and give them time and budget to do it, which... is not what happens in AAA industry these days. We'll likely be on 6k and 7k series before game devs/engines will actually pick up any of that stuff for good use. This is why we see so little FPS improvement in games. This same issue hit AMD in the past a lot too, they added fancy hardware acceleration of sound and other things and... games just did not implement it.
Rubin is the true next gen whilst NV just threw more wattage and cores at 50 series.
Just not really. It's still same CUDA cores it seems, just with even more fancy things added in hardware that AGAIN will not be implemented in engines and games for years to come, so we'll see little improvement in games most likely. In AI though... that could be much nicer. But that's not gaming, and gamers should be opposing it IMHO.
All these DLSS 5 demos now on youtube do not show VRAM being hampered, this is an architecture generation ahead of what's possible on current hardware.
How would you see and measure vRAM performance affecting the model execution? None of the youtubers I've seen done it, you need proper tools for it usually.
But it's possible this isn't really going to work well on current hardware (5090 might if it fits in the cache, CBA to play with it currently, I have other projects I am working on atm.
), or that it needs full implementation in game engine and proper masking first. Only time will tell for sure.
Remember back to Ray Reconstruction, enabling it would have on average a 15fps hit to performance, now with the latest version which has at least 25% more data to work with has no impact in framerate and in some cases the fps goes up a couple of digits depending on the engine/config and RR 4.5 is doing both upscaling and denoising at the same time now whereas it was just denoising previously -This boost applies to all RTX cards too not just the 40+.
Models get refined over time, sometimes smaller can give better results if better targeted or they use different approach. There's a lot of science involved in it, with full maths equations, not just silly AI training like many people imagine. NVidia does have full real science team and publish actual proper papers about computer graphics, but sadly often do not actually use their own research or optimise it to lowest denominator for some reason (like it seems to be the case with DLSS5 atm). Also, executing denoise and upscaling, by just logic alone, should be way faster if done in single pass by single model and not split apart into few different ones - so no wonder this works faster.
