(unconfirmed) RTX 2060 leak - 1070Ti performance at $349

[FordGT90Concept]

I highly doubt DXR is a priority for AMD. Like Gameworks, it's something no one is asking for. Rasterization will dominate for the next several generations.

 

[Totally]

Tell that to the Vulkan crowd who stood up on a soapbox pleading to let time run its course.....who are also the same people not giving RTX time to develop and dont seem to understand that cards need to be in the market for RT to happen in the first place. The irony isnt lost at all.

The card has barely been out for 3 months. Buying just for RT/TC isnt a great idea, but there are over a dozen titles coming out this year, some AAA, so we will see more of it this year as requested. How its implemented, looks, and performs, who knows, but there should be several major titles out here in early 2019 hopefully adding value to the cards.

https://www.digitaltrends.com/computing/games-support-nvidia-ray-tracing/

Oh, I did when I wasn't too busy laughing those naïve people. This happens every single time something with potential comes out and there are hopefuls hoping to be the first on that wave be it physx, mantle, CUDA were going to revolutionize gaming. CUDA did take off but not in the way gamers expected and the others got or their core feature rolled into DX and then promptly forgotten about. Now it's happening again or happened with VR, with RTX aka Raytracing 2.0 because let's not forget about Larrabee. These AMD/NVIDIA have a horrible track about releasing new tech that's going to be the next big thing and then having it fall flat. The reason I laugh because it is easy to see whether something is the future with a simple y/n question which is "Is it useful?" hence why CUDA was the only thing successful to come from any of those camps IMO.

 

[moproblems99]

The ever-increasing cost per generation is getting pretty lame.
I wonder how AMD will finally do RT... should be interesting. One thing for sure is they really need to quit lollygagging with their GPUs. Hopefully they do something with double Nvidia's RT cores and keep the none-RT performance at a tolerable level, thus giving us about 25-50% more performance than Nvidia's current RTRT performance--do that math on that one, lol (just guestimating here).

If I was AMD, I wouldn't worry about RTRT until consoles can do it. The resources they would need to devote would be astronomical for what appears to be small use case for a while. That is, unless GCN is magically good at it.

 

[FordGT90Concept]

That is, unless GCN is magically good at it.

It's not. The only reason why Turing can fake it is because of its tensor cores. It runs 1000th the number of rays non-faked raytracing requires by using the tensor cores to make an educated guess what the space in between should be. GCN doesn't have tensor cores and it never will. I doubt Navi does either so it's going to be many generations before AMD adds it.

 

[moproblems99]

It's not. The only reason why Turing can fake it is because of its tensor cores. It runs 1000th the number of rays non-faked raytracing requires by using the tensor cores to make an educated guess what the space in between should be. GCN doesn't have tensor cores and it never will. I doubt Navi does either so it's going to be many generations before AMD adds it.

My magic was simply referring to all the unused horsepower they seem to have sitting there. If they could do it, they would.

 

[cucker tarlson]

The ever-increasing cost per generation is getting pretty lame.
I wonder how AMD will finally do RT... should be interesting. One thing for sure is they really need to quit lollygagging with their GPUs. Hopefully they do something with double Nvidia's RT cores and keep the none-RT performance at a tolerable level, thus giving us about 25-50% more performance than Nvidia's current RTRT performance--do that math on that one, lol (just guestimating here).

problem is power efficiency of their architecture, nvidia can make a humongous die at 12nm draw less than Vega 64, there's not much room for AMD to deliver better than nvidia performance in RT.

My magic was simply referring to all the unused horsepower they seem to have sitting there. If they could do it, they would.

what unused horsepower are you referring to ? unless they can devise a specific way to accelerate RT,their raw Tflops number means very little. 10 gigarays is 100tflops of RT performance.Look at Pascal in the RT demo.Sure, Vega may hit 5 fps where Pascal has 4.Even 2060 would deliver 3-4x the RT performance of your fastest non-RT card like TXp.

 

[Deleted member 67555]

I have a 1070ti and I think it's nice to see this level of performance be upper mid level and I can't wait to see this level at the $200 price range where it belongs.

 

[FordGT90Concept]

As far as I can tell, there is no GPU real time raytracing benchmark that's hardware agnostic. All implementations seem to use Radeon Rays/Fire Rays for AMD cards or CUDA/RTX for NVIDIA cards.

 

[LFaWolf]

It's not. The only reason why Turing can fake it is because of its tensor cores. It runs 1000th the number of rays non-faked raytracing requires by using the tensor cores to make an educated guess what the space in between should be. GCN doesn't have tensor cores and it never will. I doubt Navi does either so it's going to be many generations before AMD adds it.

This is the first time I heard of this faking. Source?

 

[Vayra86]

I highly doubt DXR is a priority for AMD. Like Gameworks, it's something no one is asking for. Rasterization will dominate for the next several generations.

They literally even stated something along those lines and with that, consoles are also stuck in non-RT land, which makes market adoption very tricky indeed.

This is the first time I heard of this faking. Source?

https://devblogs.nvidia.com/nvidia-turing-architecture-in-depth/

 

[FordGT90Concept]

Non-faked realtime raytracing requires 1000+ gigarays per second. RTX 2080 Ti manages about 10 + denoising.

 

[LFaWolf]

I am not an expert on ray tracing but I have a hard time drawing the same conclusion as their ray trace are faked using tensor cores based on that paragraph

 

[Vya Domus]

Faked is not the right word. Imagine a shadow cast by an object that is ray traced. If the numbers of rays is low that shadow will look noisy, it will look something like this :

After this pass they simply try and interpolate the missing gaps using Tensor Cores. It's not fake, the result is simply not great. For example it will be nearly impossible to obtain sharp overlapping shadows like in the first picture doing it like this. If you noticed in all of their demos everything has this diffuse look, that's not by choice.

 

[LFaWolf]

Faked is not the right word. Imagine a shadow cast by an object that is ray traced. If the numbers of rays is low that shadow will look noisy, it will look something like this :

After this pass they simply try and interpolate the missing gaps using Tensor Cores. It's not fake, the result is simply not great. For example it will be nearly impossible to obtain sharp shadows like in the first picture doing it like this. If you noticed in all of their demos everything has this diffuse look, that's not by choice.

Now that is a much better explanation. Thank you! Spreading it as fake is like FUD.

 

[moproblems99]

Faked is not the right word. Imagine a shadow cast by an object that is ray traced. If the numbers of rays is low that shadow will look noisy, it will look something like this :

After this pass they simply try and interpolate the missing gaps using Tensor Cores. It's not fake, the result is simply not great. For example it will be nearly impossible to obtain sharp overlapping shadows like in the first picture doing it like this. If you noticed in all of their demos everything has this diffuse look, that's not by choice.

Perfect explanation!

 

[FordGT90Concept]

That's oversimplification of it.
https://developer.download.nvidia.c...PMbSVCgkJwil4mB-lNfmGChdVHTErxQfAx3RYEtrQRYGg

Jump to 20 seconds in: that's what is actually being RTRT (1spp noisy = 1 sample per pixel ray trace). For all of the lazy bastards, here's a screen cap:

Spoiler

Denoiser mops up that mess so it's more presentable (mimics combination of RTRT and rasterization).

RTX does the absolute minimum of ray tracing, ergo, faked.

 

[Vayra86]

Now that is a much better explanation. Thank you! Spreading it as fake is like FUD.

Well, not so much if you place RTRT in the perspective Nvidia likes to put it in: 'we get photorealism in graphics'. See, the current tech is also faking it, and it approximates reality pretty closely, but depends a lot on how much time and effort a developer puts into creating a near-reality.

Shadow maps for example, already cast their shadow rather sharply depending on the quality of the map. And many times, they're much more efficient at doing that level of quality than RTRT is at this point. An RT approach would probably be much more blurry. But turn the ingame shadow quality slider down, and what you get is pretty similar to that RTRT result with denoising. Its not the real thing. It tries to get as close as it can get. Not much different from existing tech, except now everything is brute forced in real time.

So, its either of two things: real or fake. I would still say that the current status is just as fake as old tech, and only the way it is implemented is just as much a deciding factor as it is with old tech as to how 'real' it seems to be.

I would much rather say: spreading it as 'real' is very misleading. Comparable to a fake Rolex, really. At first glance it looks nice, and the longer you look at it, the more tiny flaws you start to notice, only to conclude it isn't much better than a plastic POS that costs a dollar.

An example of non-RT dynamic shadows, can simply be screenshotted in a wide variety of games. You tell me... Take note of the godray accuracy and where shadows are cast on the character. Note also how light is partially cast through leaves and all other thin foliage. And this is a 120 fps, 1080p ultra shot on a GTX 1080...

And just for shits and giggles, let's look at Unreal 1 (1999). Take special note of the guy on the floor.

 

[FordGT90Concept]

Let's be honest: the only reason why raytracing makes sense in gaming is because it cheapens the cost of making games (literally transitions the burden of lighting to hardware). If you have any reason, what-so-ever, to implement a rasterization code path, raytracing has no advantage--it's redundant and requires additional debugging and level design work.

It's not going to become mainstream unless there's a landmark movement like DirectX 9 to DirectX 10 where DirectX 10 is not backwards compatible with DirectX 9 and 9 is only present for backwards compatibility sake. Games would target the new API that depreciates rasterization. The game would not work on older graphics cards that do not support the new API.

Microsoft isn't going to even consider that unless even the lowliest of integrated graphics can handle raytracing. It's a long way off. Advantage: ultra settings on the new API could look almost photorealistic. Disadvantage: "your mileage may very" has never been more true in terms of computer graphics.


I'm still surprised they're marketing 2060 as an RTX card. It's really not suitable for it, especially with a pitiful 3 GiB VRAM.