Maximum performance setup

[sergei_falaleyev]

I am building a new PC and would love to get two 5090s but so far, having no luck with my two GPUs.

I'm using Vantage 2.6.2 with RTX 4090 + RTX 3090 and here are my results of a single frame render:

RTX 3090: 0:35s
RTX 4090: 0:18s
RTX 3090 + RTX4090: 1:25s

As you can see, dual GPU setup is making the render longer rather than shorter. Is anyone else having better results?
I'm still utilizing the other GPU's 24Gb of VRAM for other applications when Vantage is working as my scenes take all lthe memory, but cannot use them both for rendering.

The developers mentioned that after 2.6.0 update the dual GPU setup works better. I did not see any difference.

[npg]

Yeah, these numbers are strange...
What is the render resolution in your test? Could you run a quick test at 720p? As discussed in other forum threads, the resolution puts a cap on the framerate because the PCI-e bus is a bottleneck between the two GPUs. This is more likely to occur if the scene is not heavy, otherwise sampling slows down the renderer enough for the PCI-e transfers to occur in the meantime.

Are you using reservoir resampling? It is not supported in multi-gpu mode. I don't expect it to have such an effect on render times, but it's better to compare renders that have the same settings in effect.

Could you try the comparison with light cache turned off? It is now supported in multi-gpu mode, but there could be some bug we didn't notice.

Is this slow speed reproducible with any scene or only with specific scenes? For example, if you use one of the sample scenes with default settings?

The specific improvements in 2.6 were:
- Multi-gpu support for Light Cache. Previously we were forced to use only brute force GI in multi-gpu mode.
- Additional render elements (if any are enabled) are not transferred to the main gpu all the time, but only at the end of the render. This reduces the effect of the PCI-e bottleneck.
- The amount of additional memory used for transfers between the GPUs was reduced.

[sergei_falaleyev]

The resolution was 1920x1600 and the scene is super heavy, taking almost all 24Gb of VRAM.

Unfortunately, I no longer can perform any more tests as I had to remove my 4090 to start some rendering work on one of my farm PCs.

I did find a few other people sharing their experience with dual GPU setup and it wasn't great either. Even developers from Chaos said the secondary GPU will speed things up 40% at most. Is that still correct? If that's the case, it would be easier building ome more PC and have it as a render slave only. What are you suggesting?

[npg]

Multi-gpu mode requires a few additional buffers. It is possible that with your scene almost filling up the memory, switching to multi-gpu could push memory usage a bit over the total VRAM. Windows/DirectX usually manages to deal with this automatically, by swapping data around, but this hurts performance.

We are aware that multi-gpu rendering doesn't scale very well. It is complicated to do for realtime rendering because we have to transfer data back to the main GPU on every pass (so that it can be displayed immediately) and this can become a bottleneck. I'm not sure about this 40% quote, there is no fixed limit. In ideal conditions I hope it can get much closer to 100%.

Note that denoising could take a significant portion of frame time and this is only done on the primary GPU which has the full frame, so the second GPU can't speed that up (though we try to balance the sampling load to compensate). For offline rendering I recommend using the "Final pass only" denoiser option. It speeds up even single GPU rendering.

[sergei_falaleyev]

Thanks for the info. Will have to do more testing.