Page 1 of 1

CUDA 10 available

Posted: Thu Sep 20, 2018 12:35 pm
by jbikker
https://developer.nvidia.com/announcing-cuda-toolkit-10

Still looking for information on ray tracing in the documentation...

Re: CUDA 10 available

Posted: Thu Sep 20, 2018 1:41 pm
by shocker_0x15
We can already find a description about new features in the next OptiX from Google search (e.g. GeometryTriangles).
According to that, the next version of OptiX becomes 5.2 :)

Re: CUDA 10 available

Posted: Thu Sep 20, 2018 3:19 pm
by friedlinguini
I don't think CUDA itself has been advertised with ray-tracing goodness; it's all delegated to OptiX, DirectX, and Vulkan.

Re: CUDA 10 available

Posted: Mon Sep 24, 2018 11:50 am
by cignox1
Honestly, I was expecting more discussion here related to the new RTX tech: game development aside, is there some interest among the offline rendering community? I suppose that even while the first iteration might not offer much in this regard (perhaps still too dependent upon rasterization?) the future plans should provide a lot of interesting features for a whole class of applications. Am I wrong?

Re: CUDA 10 available

Posted: Mon Sep 24, 2018 9:09 pm
by stefan
I would think the RTX cores are exposed the same way as the rasterisation hardware - through graphics APIs, but not general compute APIs. Note how there's no access to ROPs from CUDA or OpenCL, I guess the same goes for RT.

Re: CUDA 10 available

Posted: Tue Sep 25, 2018 3:20 pm
by jbikker
It looks like we'll only be able to access the ray tracing hardware via OptiX / Vulcan / DX12. No CUDA. If your renderer is implemented in CUDA, you can call OptiX from CUDA, but you don't get to use the hardware directly. Add to that that this is a feature promised for Optix 5.2, which is not available yet... It's "not easy" (as in: impossible) at this time to experiment with the shiny new hardware, sadly.

Re: CUDA 10 available

Posted: Thu Sep 27, 2018 10:26 pm
by szellmann
stefan wrote:Note how there's no access to ROPs from CUDA or OpenCL, I guess the same goes for RT.


In CUDA 10 there are special warp matrix operations that leverage the tensor cores on compute capability 7 GPUs (https://docs.nvidia.com/cuda/cuda-c-pro ... .html#wmma), so in general NV seem open to expose fixed-function HW through CUDA (before CUDA 10 tensor cores could however only be used via cuBlas).