Page 1 of 1

GPU Choice

PostPosted: Sat Nov 21, 2020 10:42 pm
by Nick Kurucsev
Hi,
any opinions of whether I should get 1x RTX3090 or 2x RTX3080 to replace 2x GTX1080ti?

Cheers

Re: GPU Choice

PostPosted: Sun Nov 22, 2020 1:11 am
by MishaEngel
Nick Kurucsev wrote:Hi,
any opinions of whether I should get 1x RTX3090 or 2x RTX3080 to replace 2x GTX1080ti?

Cheers


When you're not in a hurry... wait till next year.
The availability of the new GPU's (RTX30xx and RX6xxx) will be a lot better.
When you shoot higher than 4k and/or use a lot of effects pick the RTX3090 because of the memory amount (24 GB for the 3090 and only 10 GB on the 3080). No news on the performance of the RX6800 XT (16 GB) in Resolve at the moment.

Re: GPU Choice

PostPosted: Sun Nov 22, 2020 3:37 am
by Jim Simon
GPU memory isn't shared, so my choice would be for the single 3090.

I also prefer CUDA over OpenCL, so the new AMD's wouldn't be my preference either.

Re: GPU Choice

PostPosted: Sun Nov 22, 2020 8:08 am
by John Griffin
Jim Simon wrote:GPU memory isn't shared.
That’s news to me as I was going to buy another 1080ti.

Re: GPU Choice

PostPosted: Sun Nov 22, 2020 9:00 am
by Nick Kurucsev
I didn't realise that either, does that mean a 2nd card is really a waste of time??

Sent from my SM-N975F using Tapatalk

Re: GPU Choice

PostPosted: Sun Nov 22, 2020 9:02 am
by John Griffin
If memory is not shared then what about CUDA cores?

Re: GPU Choice

PostPosted: Sun Nov 22, 2020 9:50 am
by Videoneth
Nick Kurucsev wrote:I didn't realise that either, does that mean a 2nd card is really a waste of time??

Sent from my SM-N975F using Tapatalk


It's not a waste of time at all (depending on what you do of course). Resolve can leverage multiple GPUs, and depending of the task, you can get great benefices.

Check the the last "DaVinci Resolve Studio - RTX 3080 & 3090 Multi-GPU Performance Scaling" test by Puget Systems.

https://www.pugetsystems.com/labs/articles/DaVinci-Resolve-Studio---RTX-3080-3090-Multi-GPU-Performance-Scaling-1904/

Re: GPU Choice

PostPosted: Sun Nov 22, 2020 9:58 am
by Videoneth
MishaEngel wrote:
Nick Kurucsev wrote:Hi,
any opinions of whether I should get 1x RTX3090 or 2x RTX3080 to replace 2x GTX1080ti?

Cheers


When you're not in a hurry... wait till next year.
The availability of the new GPU's (RTX30xx and RX6xxx) will be a lot better.
When you shoot higher than 4k and/or use a lot of effects pick the RTX3090 because of the memory amount (24 GB for the 3090 and only 10 GB on the 3080). No news on the performance of the RX6800 XT (16 GB) in Resolve at the moment.


Add to that the fact that Nvidia will be releasing a 3070 Ti and a 3060 with more VRAM than what they put it their actual 3080 (and it's a shame, but it's good for people who were patient)

Re: GPU Choice

PostPosted: Sun Nov 22, 2020 4:57 pm
by Jim Simon
This thread provides some useful data.

viewtopic.php?f=21&t=126252&p=689043&hilit=GPU#p689043

Re: GPU Choice

PostPosted: Sun Nov 22, 2020 4:58 pm
by Jim Simon
John Griffin wrote:If memory is not shared then what about CUDA cores?

My understanding is that no resources are 'shared', and any lesser cards will inhibit better cards.

Re: GPU Choice

PostPosted: Sun Nov 22, 2020 6:00 pm
by John Griffin
Jim Simon wrote:
John Griffin wrote:If memory is not shared then what about CUDA cores?

My understanding is that no resources are 'shared', and any lesser cards will inhibit better cards.

Not talking here about mixed cards but dual identical cards. What explains the Puget systems test results where double or triple cards show performance gains?

Re: GPU Choice

PostPosted: Sun Nov 22, 2020 6:17 pm
by Andreas Schwarz
I will buy a 3070 or 3060ti in January...and will replace it later with the 3080 ti.....

Re: GPU Choice

PostPosted: Sun Nov 22, 2020 7:35 pm
by Jack Fairley
John Griffin wrote:If memory is not shared then what about CUDA cores?

It doesn't scale perfectly, but they do help. At this point, another 1080 Ti is a waste of time and money, imo.

Re: GPU Choice

PostPosted: Sun Nov 22, 2020 9:58 pm
by Jim Simon
John Griffin wrote:What explains the Puget systems test results where double or triple cards show performance gains?

That answer may require some engineering knowledge of Resolve that I don't have.

But a couple possibilities I can think of are sending different frames to each card to be processed simultaneously rather than sequentially. Or sending different processes, like NR to one and Color to another.

Could also be that my understanding is in error.