Server hardware choice for optix raytracing

Is there detailed info comparing the NVDIA hardware in regards to raytracing functionality?
We are looking into using cloud based compute and find it hard to decide which of the hardware options are best suited for our needs.

Basically we need to select server GPUs for two things:

  • raytracing with optix
  • machine learning (tensorflow)

I’m aware that there is special hardware acceleration for raytracing on the RTX series but how does this compare to the features of the server hardware like the v100 for example?

Boards with RT cores (every product with RTX in the name) will be a lot faster when it comes to ray-triangle intersections and BVH traversal.

If your requirements are ray tracing with OptiX and deep learning, then Turing or Ampere RTX boards will be the best options, since those are containing the first resp. second generation of RT cores and the second resp. third generations of Tensor cores, they will also handle deep learning tasks.

The V100 is using the Volta GPU architecture and while that is one of the fastest GPUs for use with double precision float operations, these are not generally necessary in ray tracing and deep learning. It’s also a good choice for anything which requires Tensor Cores, but Turing and Ampere boards have newer generations of those already.

Similar with the A100 data center product. It doesn’t have RT cores and is mainly architected for compute and deep learning tasks. It’s a compute monster with the largest and fastest VRAM (up to 80 GB HBM2).

Depending on your intended workloads, it’s then a question of how much VRAM you require to handle your scenes and deep learning networks.

Means for a desktop system the graphics workstation choices would start at the highest end with the Ampere based RTX A6000 48 GB board and then the smaller variants RTX A5000, RTX A4500, etc.
https://www.nvidia.com/en-us/design-visualization/desktop-graphics/

There exist data center products of those RTX workstation boards, listed in the Products column at the bottom of this site: https://www.nvidia.com/en-us/data-center/
For example the highest end data center board with RT cores would be the A40 which is similar to the RTX A6000, and so on.

Please have a look at the resp. specification sheets which list the different hardware configurations.

That site also has a link to NVIDIA RTX in the Cloud at the bottom, which goes in to more detail with available cloud based GPU setups.
Mind that data center drivers are released on a different schedule than consumer or workstation display drivers and you’d need to check which display driver versions are offered by the cloud service providers to determine which OptiX 7 minor version is supported on them.

2 Likes

Thanks for another great answer! This helped a lot. You are doing an amazing job supporting the developers here!

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.