Page 1 of 1

Benchmarks for U4K System Efficacy; input on SLI GPU use

PostPosted: Fri Jan 15, 2021 7:15 pm
by KOSTUDIOS
Hello Davinci Forum Gurus!

After working w Resolve 16 last Quarter and trying to get my build to run well in real time, I realized that the benchmarks and results I got from one build to the next for my PC system would probably help a lot of others figure out what the true min requirements to have a professional U4K video rig actually would look like...

I post some more on that topic soon, as I’m about to upgrade my rig again to the max (available) CPU for the 1176 socket for Intel, the i9-9000K ... when I do that, along with a new GPU or 2, I’ll post each stage for the forum and benefit of all, and describe what I was able to get Davinci to do, or not, at each level, and general notes as well....

In the meantime, as I’m looking to narrow down my GPU choices for PCIe 3.0 and maxing this particular motherboard out (MSI Z370 SLI PLUS ) ...

I was wondering if anyone on the forum might have a list of SLI Ready nVidia GPU cards that have been demonstrated to provide SLI multiplier advantages with Davinci Resolve on Windows 10?

Since I cant use a PCIe 4.0 card, my options are not as wild as what some of the new cards offer ( 10k cuda in one of the most recent nvidia cards, 1 GPU!) ...

But I’m hoping that some folks have been able to get SLI to work in their systems with either 30- or 20- or Titan series cards on PCIe 3.0...

And IF so... please ... share what model cards you used and what sort of benchmark improvements you saw - and Davinci results in particular, if you can...

With appreciation,
-MAVRICK @ KO STUDIOS

Re: Benchmarks for U4K System Efficacy; input on SLI GPU use

PostPosted: Fri Jan 15, 2021 8:10 pm
by richards
A PCIe 4.0 card will work fine in PCIe 3.0 system, just not at its maximum performance - but it'll be the same performance as any other PCIe 3.0 card that you could use, so it's really not a factor IMO.

If what you're concerned about is memory, Nvidia just announced a 12GB 3060 - to be released in Feb. The current 3060 Ti, 3070, and 3080 are 8/8/10 respectively, putting the new 3060 second only to the 3090 (24GB) for memory. Of course, you're taking a massive hit on Cuda cores going with a 3060 (and the 12GB has ~26% fewer the Cuda cores than the 8GB model) over a 3090, so even in SLI you're not going to be close. I'm sure we'll see new 3070 and 3080 announcements in the coming months though, but who knows what they'll bring.

As was clarified by BMD in a recent thread here, in SLI memory is not additive (it's the minimum of all cards), so you have to go for cards with as large memory as possible. Cuda cores are additive in SLI, except you are taking the not-insignificant performance hit of the card-to-card copying and communication.

That makes the decision pretty easy I think - if you want both maximum memory and Cuda cores, you have to go with the 3090 currently. It'll be significantly better to have a single 3090 than SLI anything else. If you can afford SLI 3090's, then that's just icing on the cake!

Re: Benchmarks for U4K System Efficacy; input on SLI GPU use

PostPosted: Fri Jan 15, 2021 8:25 pm
by roger.magnusson
Recommendation from BMD has always been not to use SLI for DaVinci Resolve and they don't support such configurations. They support multiple cards without SLI but it's only useful if each frame requires a lot of GPU processing time.