What hardware for extremely large CAD assemblies?

[Anonymous]

I work with very large CAD assemblies, and my current machine (i9 7900x, 2x P5000 SLI) just isn't up for the task. I have to spend considerable time aggressively merging nodes to avoid bottlenecking the CPU as well as finding the lowest tessellation quality that works without artifacting to avoid bottlenecking the GPU.

The assemblies can have as many as 40,000 nodes after import and initial cleanup.  Polycounts could easily be 300M+. I am trying to put together a workstation spec that would allow me run as many nodes and polys as possible. I am only interested in running the maximum possible geometric fidelity with the most possible nodes, I don't use any cosmetics like textures, shaders, or complex materials.

CPU:

How much will increasing clock speed increase the number of nodes I could conceivably run? Would going with a high clock/low core processor such as the Xeon W-2125 + liquid cooling and aggressive overclocking yield any meaningful increase in feasible node count, or is there a number of nodes that VRED isn't generally capable of exceeding?

GPU:

With all the new developments in the last few months, there's lots of questions here:

- Will VRED support/see any benefit from NVLink over SLI?

- Will VRED be able to take advantage of more than two graphics cards via NVLink or other methods?

- Should I be looking at Titans or high-end Quadros?

- Does VRED see any benefit the Tensor cores in Volta and Tesla GPUs?

- Would Volta (Titan V/GV100) or Tesla (GTX2080/GTX8000) be likely to perform better for this application?

Other:

- I've seen research that points to disk I/O as a hidden bottleneck in some high-end VR applications, with non-trivial framerate increases achieved by moving to software-raided m.2 SSDs. Would this be worth doing with VRED, or will disk I/O likely not be a problem?

Thanks in advance!

[michael_nikelsky]

The number of nodes is mostly only relevant for OpenGL. No hardware will solve that issue since you will be CPU limited due the pure number of drawcalls. For working with such large scenes I would recommend to use raytracing instead and get a machine with as many cores as you can afford. Raytracing is much less sensitive to the number of triangles or nodes in the scene. 

Still, scenegraph based operations will still become a bottleneck if you have too many nodes so merging it still beneficial.

Kind regards

Michael

[michael_nikelsky]

An additional recommendation: For OpenGL you should try to turn on occlusion culling under Visualization->Advanced OpenGL Settings->Occlusion Culling. This will often massively reduce the number of nodes rendered and should give you a decent speedup.

[Anonymous]

Sorry, I should have mentioned that I'm viewing all this in VR not rendering static scenes, so I'm pretty sure I'm stuck with OpenGL, unless you think I could get raytracing to work in stereo in real time. And I am using occlusion culling, even with it on I'm still having to aggressively optimize my assemblies. It seems like my current limit is around 3000-5000 nodes before the CPU starts to lose its ability to keep up with the GPU, am I understanding correctly that this is unlikely to improve even with significantly faster clock speeds? We use VisionaryRender with the Vive for some things here as well, and its able to function with a much larger number of nodes, so I would think there would be some sort of solution to this bottleneck. Having the ability to interact with a full assembly on a part-by-part basis would be hugely valuable for us.

Are you able to comment on the efficacy of the new tensor cores, or potential performance differences between Tesla/Volta cards  in VRED, or any of my other GPU questions?

[Anonymous]

Michael,

Would you at least be able to comment as to whether VRED works with NVLink, or NVlink with >2 GPUs? Or whether I'd see better results with high poly counts using Titans/2080s vs Quadros?

Thanks!