NVIDIA's Work On Adding Ray-Tracing To Vulkan

[oleid]

1. There's no Vulkan<->OpenCL interop.
2. Drivers (especially on AMD) are all over the place.
3. I think OpenGL<->OpenCL interop is pretty inefficient.

It would seem they have some sort of interoperability at AMD:

https://youtu.be/xyP3Ut97zVY?t=382

[tmpdir]

Although it's interesting, I'm just getting out of this article that vulkan is already positioned to follow directX, like OpenGL did

[VikingGe]

3. Not necessarily in most cases, typical use case would be fluid dynamics simulation, where CL would compute "waves" and GL would render them using GPU shared memory (eg. CL would modify vertex buffers for GL)

And why would that be preferable over using GL compute shaders when you'd use them in a Vulkan version anyway? I highly doubt that the CL approach that has to go through a different state tracker would be nearly as efficient as just using compute shader.

[artivision]

This renwed interest in Ray tracing comes too late , neural networks can already be used for ilumination , and soon enough will be capable of reliably hallucinating high quality video in real time

Neural_Networks and DSP/Tensor_Cores the maximum they can do is to run a De-noising filter so that way the real Shader_Cores can calculate fewer rays, but than can be done by an ASIC as well.

[Dedale]

That reminds me of the good old days of povray. Installed it on xubuntu but is unusable. Project seems half-dead and not widely used on linux.

[artivision]

Basically "Tensor core" is a glorified FPU, it can't decode instructions, it can't execute instructions.

Here is rough GPU vendor marketing translation with CPU (near) equivalents:

SM/CU -> CPU core + L1 cache + stuff
TMU -> Load/Store pipeline
Cuda Core/Streaming Processor -> 1x Lane of vector/SIMD execution unit (FMA*)
Tensor core -> Accelerated "Cuda core" that works on 4x4 matrix

*FMA - fused multiply add, aka "a*b+c" on matrices (weights[i]*inputs[i]+biases[i] in neural networks)

https://youtu.be/KHa-OSrZPGo (GPU vs CPU hw)




Reusing known and working code that you got somewhere is almost always preferable to rewriting it from scratch, especially in the example case that I used before.
Or if you wanna support < 4.2 GL revision (or whenever was compute shader added, don't recall exactly) for example.
Those middleware things were often written for CL, especially for simulations where realtime performance doesn't matter since you are not trying to hit next frame under 16ms (or 6.9ms).

Nvidia's Tensor_Core is 640*4x4x4*FMA(mixed_32b<-16b*16b+32b), It isn't 4x4, it's 4x4x4. The other known definition for all Tensor_Core technologies is DSP, it's just what they are. Also they cannot calculate Rays or anything similarly complex, they may use them for a De-noise filter as i wrote above or we may never see them on Gamer GPUs, don't wait for them it's futile.

[miabrahams]

This renwed interest in Ray tracing comes too late , neural networks can already be used for ilumination , and soon enough will be capable of reliably hallucinating high quality video in real time

This is the kind of outlandish thinking that makes people shell out thousands for a useless "self driving" upgrade for their Tesla that gets canceled a year later. Repeat after me: neural networks are not magical.

[vein]

One that is based on OpenCL and therefore not usable for games. Or did I miss something here?

I think Nvidia just want to wire up their RTX stuff which they are already using as a DXR backend to Vulkan. Nothing wrong with that, although I have my doubts that anybody is actually going to use it any time soon.

Well, it seems that Unity 3D does not agree with you:

https://blogs.unity3d.com/2018/03/29/amd-radeon-rays-integrated-into-unitys-gpu-progressive-lightmapper/?sf185787736=1

I am very much looking forward to 2018.2 release with a preview of this functionality.

[VikingGe]

Well, it seems that Unity 3D does not agree with you:

They explicitly mention that they are going to use it to bake light maps, not for real-time rendering which the current Dx12/Vulkan raytracing hype is all about.

[vein]

They explicitly mention that they are going to use it to bake light maps, not for real-time rendering which the current Dx12/Vulkan raytracing hype is all about.

They do? Maybe I did read it too fast...but I did a search on the page for "real-time" and only found (Quote):
"The Real-time Ray Tracing with GPU Progressive Lightmapper is expected to be released later this year."

I'll read it again when I got the time (at work now) so tha tI can be sure.