Adshir's LocalRay mobile-friendly realtime raytracing solution *spawn

I do wonder if its more about the fact that RTX isn't able to use the data structures it uses, but maybe AMD could. He wouldn't be in a position to say. Due to nda's
Wasn't sure if it's his English or not, but he could've easily said upcoming hardware without mentioning amd. So I'm not convinced that's the case. But is possible.

 

There's always Google Stadia... But not sure why they wouldn't just upgrade video card hardware instead once AMD sells it.

 

I don't think having accelerated RT means you wouldn't still look at other options for specific cases, too? I mean, they said they have deal with one of the console manufacturers, but that doesn't mean it would be the only RT solution used by said manufacturer.

 

Yeah, maybe. MS has talked about 'intersection performance', not the usual GRays/s.
So this could mean it is possible to have your custom acceleration structure, as have been said in the forum. Triangle and box intersections are utilized in any case.
It could also imply there is no BVH traversal HW at all, because if so it would always beat custom software i guess? Either this or software solution is just more effective in certain special cases.
(damn... Cerny just talked, but still speculating...)

 

Cerny at least confirmed PS5 uses the method described in AMDs RT patent (but again, that doesn't mean they couldn't use LocalRay too for whatever specific needs it might suit better)

 

He only confirmed to use AMDs RT solution, but we don't know how close it is to the patent. Or did i miss something more specific he has said?

 

Possibly the part where he mentioned intersection engine in CU (TMU is part of CU, which is where the patent places the intersection hardware) and that shader sends the instructions to check this rays interesections to said engine (described in detail in the patent) and is free to do other things while the intersection engine does its job

 

At least, so far there is nothing indicating AMD would do any different from the patent.
I was not really sure if the TMU also does traversal logic or if that's entirely part of programmable CU, though. I remember they used BVH4. I assumed that's hardware standard but could be also just an arbitary choice of the programmer.

Still, no traversal shader support is confirmed, also nothing mentioned to further extend the upcoming Vulkan RT API which so far only conforms DXR feature set.
But we'll see. I'm optimistic for more flexibility, and Adshir's claims would make sense then if they are related to SX / PS5.

 

So Sony are using local rays in order speed-up their ray tracing calculations. And if they are using Adshir's LocalRay tech (they already announced a deal with a console manufacturer and they are even using Spider-man to showcase their tech), then Sony wouldn't even need the denoising part, so no need for tensor cores.

 

How do you get to the conclusion they would use any third party SW / HW? (well, Spiderman is a hint...) Why 'local rays'?

No ML denoising got out of research or offline application yet. All games do it with compute. (I expect NV might have something with Ampere launch to strengthen their point about tensors in gaming GPUs, but we'll see...)
But why do you think Adshir has an alternative to denoising? The robot demo was about sharp reflections only, and there is no need for denoising.

I'm puzzled - what could Adshir offer on top of AMD HW RT and a typical RT API that is interesting for Sony?
I could only imagine AMD is indeed more flexible allowing programmable traversal, and Adshir has some nice SW library to utilize this. But DX12U does not show anything of this, so i'm not that optimistic.

 

I could see Sony licensing to provide as part of there sdk.
Just like other tools and apis they provide.

I don't know if it provides anything that Sony couldn't do themselves, but a lot of the time that's besides the point if you can save time and money licensing it.

 

Local ray tracing just fits well with how Sony describe their tech. I am not saying they are using Adshir API.

But yes about the denoising part, they say there is no noise, not that they don't need to denoise.

 

That image and text is pretty much as general description as possible.

I would love to see the presentation about this.
Their demos a while back had all the usual aliasing problems of tracing against polygons.

If they meant that with mirror like reflections or for point light shadows they do not need denoising, yes in those cases standard AA methods do work.
For area lights/shadows fuzzy reflections any path tracing variant, denoising becomes quite good idea.