2x 2080 TI Blacks 11GB each with NVLink on PCIE V3.0 vs. 1X 3090 TI on PCIE V4.0 Performance Questions

[jujubee]

Wondering what you would think performs better overall:

1, Running 2x 2080 TI Blacks 11GB each with NVLink on PCIE V3.0 board which causes cards to drop to x8 lanes each. (Intel Z390)
2. Running 1x 3090 on PCIE V4.0 board which is a x16 lane. (AM4 board)


I would think the 3090 would be advantageous on the PCIE v4 board. The pcie v4 board also supports faster RAM, NVMe Drives, etc. but maybe I am wrong and the performance is negligible.

3. Also, how would 2x 3090s with NVLink perform on the older V3.0 motherboard running at x8 each compared to #2 above?


I ask because I'm buying new hardware and didn't want to spend more money on new motherboards and processors. Thanks!

[jujubee]

Update. I didn’t want to open a bunch of 3080s and 3090s but I’m running a bunch of benchmarks right now on them to find a price to performance ratio.
I’ll also run tests on a PCIE 4.0 board (single GPU only) to compare how pcie 4 devices (NVMe, better ram, etc) makes any substantial difference.

Will post the results likely tomorrow.
The problem in finding with vray benchmark is that it doesn’t go into system specs, only GPU(s) being used and OS. It would be nice to see chipset, onboard RAM, etc. I’d also like to know if people are overclocking because it is something I refuse to do due to instability.. I don’t see any windows 11 benchmarks either.

[vladimir.nedev]

You can see some performance tests we did here
https://forums.chaos.com/forum/chaos...3080-on-lavina

A single 3090 is definitely a better choice compared to 2 x 2080 Ti.


Greetings,
Vladimir Nedev

[npg]

Note that the PCI-e generation and lane count (or NVLink) only really matters to Vantage if you're using two GPUs because they will be using the bus a lot, all the time (rendered pixels need to be gathered on the "main" GPU). A single GPU will only use the bus when uploading geometry and textures so it could only have a minor effect on initial load times.