NVS 810 for multiple outputs

Hello,
I have a project which i would like to have a large number of displays playing simple content togather for. I am thinking about using a system with a couple of NVS810 graphics cards. Are there any bottlenecks i should be considering? Are the NVS series cards able to take advantage of GPU affinity? Thanks for your help.

The NVS 810 is actually two GPUs, so to address all 8 outputs you do need to use GPU affinity. One process can only address 4 of the outputs directly, unless you use Mosaic, but that will result in worse performance.
It is a very low power card, so it’s ok for video playback, but not much else

Interesting Malcolm, when you say “One process can only address 4 of the outputs directly” , does this mean that across the board any TD instance can only “see” a single GPU? I’m considering building a 2*P5000 system for 8 outs from a single TD instance, will the second card be inaccessible? I’m aware mosaic can span gpu’s, which presumably would let TD span all outputs, but then you say that performance would be much worse. This is contrary to my experience of mosaic being the only way to have smooth performance, I don’t understand.

FWIW, I think it’s probably worth updating the docs on these types of issues considering their importance. The docs have old information like “We have not tested to see what benefits are gained from using this configuration” and no info on windows 10. I also see paid training material from third parties on this topic citing nvidia and derivative as the source of information, which makes me wonder why this info is under lock and key…we just need to output our work. Are there any plans to refresh the info we need on this topic?

Yes, a single instance of TD can only see a single GPU (unless GPU affinity is used). If you have two GPUs in a system the driver will make one of the GPUs do all the work, and if the other GPU needs to display content the final image is copied over to it for facilitate that. Only GPU affinity will allow TD to see different GPUs.

When I say the performance will be worse, I mean the performance would be worse with 2 GPUs using SLI Mosaic, than if you were able to solve the same issue with 1 GPU. For example using a DataPath splitter with 1 GPU for 8 outputs will perform better than using 2 GPUs for 8 outputs. Of course that’s not always possible, but it’s just a warning that using 2 GPUs in mosaic mode does not add power, but instead actually costs some.

The “We have not tested to see what benefits are gained from using this configuration” note is related to SLI, which remains true.

I’ll see what I can update on the help, we certainly don’t keep any information under lock and key on purpose.