Hello,
Does anybody have an idea if the Tensor Cores available in RTX graphics cards help speed up GPU simulations (like turbulenceFD or Realflow Dyverso solver) ? and what about Optix Denoiser found in Redshift ? (in comparison to non RTX GPUs).
All this “AI stuff” is still blurry to me, all i find is about Deep learning topics but nothing related to DCC software.
Regarding Optix Denoising with RTX vs. non-RTX: most NVIDIA-cards of recent years have Tensor-Cores. The RTX-cores are not used for AI. And Optix existed before RTX. So it should have no impact here. As far as I know RTX-cores do only one thing: hardware ray-triangle intersection. Maybe they also speed up bounding volume hierarchy creation, but I don’t think so. This also means they are not usable for volume rendering.
This also applies to the fluid question: this is also possible with pre-RTX-cards. A 1080 should be fine for that.
I followed AI-based fluids only a bit. But I am pretty sure it will come. Just look at the amount of AI papers at SIGGRAPH this year… Also stuff like complex deformation or creation of walk cycles is heavily researched and I wouldn’t be surprised to see something like this pop up soon in 3D packages. In 2D this is already reality, Photoshop comes with a lot of AI arteady.
Interesting…
Based on this Nvidia Blog post about RTX vs GTX:
It seems like those “Tensor Cores” were only added recently with the RTX GPUs just like the “RT cores” ones.
quoting from that post:
“RTX graphics cards are also the first to offer Tensor Cores capable of delivering over 100 teraflops of AI processing to accelerate gaming performance with NVIDIA DLSS.”
followed by a detailed comparison of 20- vs 16- vs 10- series
…
|Ray-Tracing Cores|Yes|No|No|
|Tensor Cores|Yes|No|No|
…
What i did understand was, now that they have “dedicated” hardware would make their specific tasks much faster in comparison to everything being mixed in the “CUDA cores” of previous generations.
“RTX also relies on a fancy new denoising module… Nvidia uses this technique to help it create frames more quickly. It is likely this capability builds on the AI-powered work Nvidia presented at SIGGRAPH last year, as performing it in real time relies heavily on the Tensor “AI” Cores in its Turing GPUs.”
source: https://www.extremetech.com/extreme/266600-nvidias-rtx-promises-real-time-ray-tracing
Ok, it seems the truth is somewhere in the middle:, they were introduced with Volta in 2017, but were only part of some Quadros and the Titan. Still, you can do machine learning on a 1080 or even my 960m pretty well, so I guess NVIDIA cards support ML even without the new tensor cores.
Of course you could do ML before Tensor Cores on a GPU the same as you could use ray tracing before RTX. It is just specialized hardware on the GPU to accelerate certain calcucations.
turbulenceFD and Realflow don’t see any benefits in Tensor Cores or RTX yet. Maybe it will come some day. It’s up to the developers.
Optix Denoiser also works on non RTX cards.