Discussion in 'Article Discussion' started by bit-tech, 22 Aug 2018.
The company's going in hard, too, with partnerships
brown envelopes stuffed with cash that have seen a wealth of games pledge support for the new features of Nvidia's Turing architecture. FTFY. I wonder how long after the release the ray tracing support will be added.
Already making excuses. RTX is off to a flying start, eh?
Does that mean no extra performance over the previous cards in current games? Surely not.
It's entirely possible the cores are more efficient in some way, which would allow them to outperform current-gen cards even with fewer physical cores - we won't know until we see benchmarks. What we *do* know is that a GeForce GTX 2080 Ti with all CUDA cores would outperform the GeForce RTX 2080 Ti with its mixture of CUDA, RT, and Tensor cores for games which do not make use of the RT or Tensor cores (which, at the moment, is all of them.)
If you normalise to the same die area (and same manufacturing cost), that is. And assuming that doesn't result in a bottleneck somewhere else (e.g. memory bandwidth to feed all those CUDA cores).
Cost has no impact here. The same die area is a given, but filled to the brim with "traditional" CUDA cores instead of the massive area reserved for ray tracing. See here, the green block is just ray tracing:
Bandwidth should not be a bottleneck, what with a 352 bit bus and 14 Gbps GDDR6.
Not sure we can make that assumption since Raytracing may or may not carry significantly higher requirements.
But yes, for 99.9% of games it won't be an issue.
Well, we were talking about theoretically replacing all RT cores with CUDA cores, so we are in known territory for bandwidth.
That's a Pretty Slide Picture rather than an actual die layout diagram.
If you look at the die shot (assuming even that is accurate) you can see 144 SMs for the fully operational die. Looking at the Quadro RTX 8000, that means 32 CUDA cores per SM (to make up 4608), 4 Tensor cores per SM (to make up 576) and an unknown number of RT cores
::EDIT:: Herped a derp, SMs are mirrored. 72 SMs, 64 CUDA cores per SM, 8 Tensor cores per SM. A little birdie also says 72 RT cores, but not whether that is 1 per SM as part of the SM or a separate block independent of the SMs (I'd guess as part of the SM myself).
If we go by previous generation die layout breakdowns, the central 'column' is composed of the main uncore controller in the centre, the 'top' and 'bottom' of the bands being the ingest pipelines. Around the periphery are the memory controllers and ROP units (combined).
"Windows, Xbox One, and PlayStation 4, with only the former receiving ray tracing support" Wouldn't this be to do with consoles not having RT cores??? Shouldn't be a suprise
Separate names with a comma.