Nvidia Turing Speculation thread [2018]

Status
Not open for further replies.
I thought Nvidia already supported double-rate fp16.
 
Wouldn’t make sense to issue FP16 instructions to tensor cores. They’re setup for much denser matrix math. The “normal” FP32 CUDA cores can do double speed FP16 just fine.

I thought Nvidia already supported double-rate fp16.

Yeah, Nvidia already supports that but they artifically limited fp16 on Geforce line. Since tensor cores in 2080ti apparently has capability to push 110Tflops on FP16, I was wondering how it can be used to assist other graphical functions other than the purpose of denoising on RT.
 
I'm trying some reverse math... This at best may be a very rough approximation. But has there been any indication of how may rays per pixel you need to be able to calculate for ray tracing to look good enough?

With 10 Gigarays per second (2080Ti) at 4K and 60 fps, this translates to 20 rays per pixel (10e9/3840/2160/60). Is that good enough? Maybe in the real world assume half that plus the denoiser...
This is 1080p at 1spp. Already looks far superior that rasterized lighting, IMO:


Also it's from a year ago and it's not even using AI.

At least for diffuse GI (irradiance) I think you could get away with tracing at 1/4 the resolution and still look good.
 
AMD needs to announce something DXR-related soon imho.

AMD needs to get their "traditional" rendering performance up to speed on PC solutions well before they go after a feature that will only be present and barely useable on $1k graphics cards and 3 games.
 
Any reports on how the actual Ray Tracing hardware is designed and configured? They are rays so I'm assuming just a super specialized form of vector SIMD? Nvidia's descriptions seems like a sort of "bolt on" measure that can be omitted for non-RT cheaper Turing based GPUs (think GTX 2060). Is the RT silicon added arbitrarily to GPU with proper comm pathways or are the RT blocks part of the SMs itself? If the latter is true, then an RT-less Turing SM I guess wouldn't necessarily be a Turing SM at all. RT on SM would make a better case for a top to bottom product stack under the RTX moniker than having a top end RTX line with a middle and low end GTX line to handle the low end. Makes for bigger GPUs that might lack the tensor cores and in turn alot of the ray tracing grunt that comes from the big Turing implementation.

I'm sure Nvidia is working on a low end RT fall back to accomodate cheaper products.
 
I'm trying some reverse math... This at best may be a very rough approximation. But has there been any indication of how may rays per pixel you need to be able to calculate for ray tracing to look good enough?

With 10 Gigarays per second (2080Ti) at 4K and 60 fps, this translates to 20 rays per pixel (10e9/3840/2160/60). Is that good enough? Maybe in the real world assume half that plus the denoiser...

Others already have pointed out that the 10 Gray/s would be for 1 on screen triangle.
In principle you have to test each triangle (also off screen triangles for secondary rays) if it intersects the ray.
With 10 million triangles, if you test each triangle, you could be down to 1 thousand rays per second.
Acceleration structures like BVH can bring down the number of intersections, but currently we have seen no information how fast the RT core can traverse a BVH and how many remaining triangle test per ray there typically remain.
Acceleration structures help reduce the number of ray/triangle tests, but they have as drawback that they have to be rebuilded when the geometry changes, like for skinning, dynamic hair/vegetation etc.
We have no information either on what is the cost of rebuilding the BVHs each frame.
 
Well, the 2070 is priced considerably lower than the 1080 Ti...
User-centric market values:
1080 Ti -> 700 EUR (settled prices)
2070 -> 640 EUR (one-vendor only price)

Yes, I am aware that the presentation spoke of 499 US-$. But that's without tax, without tariffs and most importantly that's for the cheapest-possible AIB models (blower fan, plastic shroud, maybe dumbed down PCB). The better AIB models won't get near that mark for a while.

Well... RT is such a step up. Having a non RT capable gpu will be like having nothing in 1-2 gpu refresh.
Then maybe that would be a good time to get a new GPU. :)
 
Last edited:
Yeah, Nvidia already supports that but they artifically limited fp16 on Geforce line. Since tensor cores in 2080ti apparently has capability to push 110Tflops on FP16, I was wondering how it can be used to assist other graphical functions other than the purpose of denoising on RT.
GPUs in GeForce line have different SMs to the SMs in P100 and V100 so it's not artificially limited there simply is no capability to do it fast. Tensor cores are "weird" in more ways then one if you want to look at them for something like this. As already said they are optimised for matrix operations meaning you cant simply pack arbitrary scalar code on to them. They are CUDA only and there is no equivalent for that in D3D (but you can interop). They require huge amounts of internal bandwidth to the point that using them naively in CUDA will result in only about half of the peek performance. Also its not completely FP16. It takes FP16 input, but the math and output is done at FP32.
 
Last edited:
I will skip this gen but it’s extremely impressive that there are at least 3 top tier titles in the works and launching soon. Hopefully it doesn’t end here.

I see a lot of griping about nvidia wasting transistors on tech that will be hardly used or that will be too slow. But what’s the other option? If we wait until general compute is fast enough for raytracing it’ll probably not happen for at least another decade.

The real kicker will be the next console generation. If there’s no support for RT acceleration in consoles (or by AMD) then it might as well be dead until 2030. I’m optimistic that raytracing is actually easier to implement than all the hacks and that there are enough raytracing fanboys in the developer community to have DXR really take off.

I cannot wait to see how much faster Nvidia's implementation of raytracing (the RTX platform) gets with future generations Nvidia GPUs on 7nm, 7nm+ with EUV, and 5nm

At the same time, I tend to agree with you about the outlook for raytracing support in games if there's no RT acceleration from AMD, and the next console generation.

So then we'd probably be looking somewhere between the late 2020s and 2030 with 10th generation consoles.

[/QUOTE]
 
User-centric market values:
1080 Ti -> 700 EUR (settled prices)
2070 -> 640 EUR (one-vendor only price)

Yes, I am aware that the presentation spoke of 499 US-$. But that's without tax, without tariffs and most importantly that's for the cheapest-possible AIB models (blower fan, plastic shroud, maybe dumbed down PCB). The better AIB models won't get near that mark for a while.


Then maybe that would be a good time to get a new GPU. :)


Oh yes, my vega FE is and will be fine for, I hope, 2 years at least ? But I hope post navi will have RT (or Intel gpu if they support freesync) hardware support.
 
Yeah, Nvidia already supports that but they artifically limited fp16 on Geforce line. Since tensor cores in 2080ti apparently has capability to push 110Tflops on FP16, I was wondering how it can be used to assist other graphical functions other than the purpose of denoising on RT.

Only GP100 and GV100 had accelerated FP16 in the past (ignoring Tegra related products), was not on any of the other GPU models from Pascal.

With the latest GPU need to consider that this is now SM_80 and brings acceleration-rapid packed maths/further optimisation to mixed precision GEMM with CUDA 10.
Also Nvidia has further enhanced this feature/compute CUDA level by supporting closer integration with DX12 and Vulkan, but it needs to be seen exactly what interoperability has been exposed, unfortunately that level of info will not come out until they present more on CUDA 10/SM_80 but it offers some potential benefits as some algorithms/features could never be accessed by gaming development in the past, and tbh some of those are more efficient than what was implemented in Gameworks.
SM_80 and associated compute capability would be Turing, so would only apply to their generation and going forward.
 
Last edited:
Not only 3 games, there are more coming.
There are more games with raytracing? I got the impression that only those few used raytracing. They announced a bunch more with the "developed for RTX" moniker, but it seems this can be as vague as just supporting deep learning super sampling.

The only title I could see as very high-profile on PC would be Battlefield V, but it looks like the game may tank due to EA/DICE's unhealthy fixation on selling coloured cosmetics on a WW2 game. DICE's CEO of 14 years was already booted, reportedly due to abysmal pre-order numbers.

Still, yes it's going to be a niche for sometime, but AMD needs to support it eventually.

I was quoting a post that suggested AMD needed to announce something with DXR "soon". They really don't.
If it actually gains traction then they'll be forced to do it eventually, sure.

IMO, in the near future AMD should attempt to take full advantage of 7nm to increase raw performance on DX11/DX12 workloads, not maintaining feature parity with the super-expensive RTX family.


Is the raytracing path on these 3 titles vendor-agnostic or is it locked to the RTX GPUs (or behind gameworks)?
 
Last edited by a moderator:
I believe it's working with DX RT (I don't remember the name), so it's not nvidia locked ?

Would be funny if Nintendo can pull a switch 2 with hardware RT (even if the performances are anemic with the need of a very small soc of course) while the "big consoles" don't have it.

I believe RT will take, because it's something visible. Async Compute, packed math, etc, it's maybe as important, but you don't "see" it for the end user, it's harder to sell to the final customer. And nVidia have a lot of ressources. It's not AMD, who can't even have their very good idea (primitive shaders) working on their own gpu...

Anyway. Right now, the main problem for nvidia is emptying the pascal stock I guess...
 
Last edited:
It seems more inconsequential relative to physics features if locked within post processing of Nvidia when one considers no-one complained about VXAO (especially how implemented in Final Fantasy 15 where it uses integration with a custom pass SAO within Luminos Engine), or even complain that much about Turf Effects where it did not cause much of a controversial stir.
Point is it becomes more of a sticking point if it was physics engine related simulation (possibly could lump Hairworks into this category) rather than the actual Ray Traced Area Shadows\Ray Traced Glossy Reflections\Ray Traced Ambient Occlusion that are in Gameworks.
But then one does not necessarily need to use those libraries depending how accessible it is in DirextX and Vulkan that are also supported and how evolve, I assume comes down to expertise support from Nvidia-Microsoft-Khronos/engine integration.

Caveat my context/point is not about performance of Gameworks or how well it is actually implemented-used by Devs (some use it better than others), raytracing features in it probably will not be perceived in the same way as some of the physics simulation libraries.
 
Last edited:
NVIDIA is listing 21 Games with RTX Support
For RTX to succeed, they need proper implementation. And it seems that NVIDIA has been working hard with many game developers to get the new technology supported. The slide shown shows 21 titles with Shadow of the Tomb Raider and Battlefield V having support at launch from what we heard.
  • Ark: Survival Evolved
  • Assetto Corsa Competizione
  • Atomic Heart (2019)
  • Battlefield V
  • Control
  • Dauntless
  • In Death
  • Enlisted
  • Final Fantasy XV
  • The Forge Arena
  • Fractured Lands
  • Hitman 2
  • Justice
  • JX3
  • Mechwarrior V: Mercenaries
  • Metro Exodus
  • PlayerUnknown’s BattleGrounds
  • Remnant from the Ashes (2019)
  • Serious Sam 4: Planet Badass
  • Shadow of the Tomb Raider
  • We Happy Few
https://www.guru3d.com/news-story/nvidia-is-listing-21-games-with-rtx-support.html

Should be just another option in advanced game setting to enhance graphics (RTX on/RTX off).
 
I’m sure folks will pay whatever for the 2080/Ti but 500+ for the 2070 is rough.
 
Status
Not open for further replies.
Back
Top