Optix performance on Nvidia Quadro RTX Server

Hello, we’re contemplating the purchase of a Quadro RTX Server for real-time ray tracing:

Are there any benchmarks showing how Optix scales in multi-GPU mode on these servers?

Hi @SuperGastrocnemius!

That is a good question, and I’m asking around internally for a better answer than I have at the moment, so I’ll update here as I get more info. I’m not sure yet exactly what network configurations RTX Servers are available in, but you’ll want to pay close attention to the PCI & NVLINK setup.

I assume your team is building your own renderer? Or are you thinking of using this server with an existing commercial GPU renderer? One suggestion I got back, which might help a little, even if you’re using an in-house renderer, is to look at the scaling benchmarks for the existing third party renderers, such as Octanebench for the Octane renderer.

When evaluating scaling with OptiX, the first question to ask is whether you need memory scaling; how much memory sharing needs to happen on input and on output, and what memory exactly needs to be shared. You can obviously get near perfect scaling for batch jobs and when rendering each frame separately on a different GPU, as long as I/O isn’t bottlenecking feeding the GPUs. If you want all GPUs working on the same frame, the next best thing might be to duplicate the scene on each GPU if you need to write to a shared output buffer. Be aware that OptiX currently doesn’t share output buffers across PCI islands, and scaling perf when sharing output buffers can be sensitive to how well accesses are aligned and localized, and of course whether you need any read/write locking. OptiX will share textures over NVLINK, but be aware that by default, OptiX duplicates texture memory until devices are mostly full, and only then will start sharing memory by assigning some texture buffers to individual GPUs, so memory scaling may not be initially visible to you. Also keep in mind that OptiX doesn’t currently share acceleration structures, they will be duplicated on each device, which may limit your memory scaling.

I would recommend that you contact optix-help or an RTX server rep so we can discuss your needs in more detail and make sure the RTX Server can be configured to adequately meet your needs.


Thanks for your reply, David. WIll get in touch with optix-help.