2020 GPU questions. 1 GPU with max vram or 2x

[Samir Patel]

Hi all, been a while since I upgraded my Nvidia cards. Currently running 2x GTX 780 cards with 6GB each of vram. I've been just making some notes regarding future upgrades. And now with recent updates to Resolve and Premiere Pro I'm a bit confused.

Wondering what would be better for going forward. More vram in one card? Dual cards because more cuda cores together? Kinda confused.

My current cards would be 2304 cores each and 6gb vram each.
Eyeing either
1x RTX 2080ti (4352 cores and 11gb)
2x RTX 2080ti (2x 4352 cores and 2x 11gb)
1x Titan RTX (4608 cores and 24gb)

Thanks hive mind. Any help would be appreciated

 

[Dermot Shane]

wait for RTX3080ti and BigNavi

not the optimal time to be investing in GPU with newer / faster / cheaper on the near horizion

 

[Samir Patel]

Hi Dermot, good point. But can you tell me still if it's better for more vram in one card or more cuda in one card or dual cards etc? May be a long while before I can upgrade my workstation so considering just doing GPU now (at least after RTX3080ti drops)

 

[Dermot Shane]

depends on how well the software in question shares GPU's
For Resolve the dev have said the best single GPU you can afford is optimum since v15
i'm running a machine with a single 2070s at home and when driveing a 4k timeline it's not far off my machine at the studio with dual 2080Ti's
there is a a substaintial diffrence when exporting a 4k master tho
any more VRAM than is needed is a total waste of resource, more is not better, more wastes power, adds heat and gives you zero in return... so really defineing where the line between "not enought" and "too much" is very dependant on workflow and delvierables
i find 8g enough for a UHD timeline with some caches in play
and 11g is enough for aUHD timeline with next to no caches in play
i deliver 4k and UHD so my timelines and monitor match the deliverables
the only wild card is haveing to deliver cam orig, and the cam is Helium
the few times that's been on the table i've sent 8k DNxHR444 back to editorial to test, and their machine melted down... so that became a non issue and i dleivered 4k that they could actualy use to finish
no idea how Pp leverages dual GPU's
i'd get the minimum needed today, and keep my resources avb for when the new tools are reality

 

[Samir Patel]

depends on how well the software in question shares GPU's
For Resolve the dev have said the best single GPU you can afford is optimum since v15
i'm running a machine with a single 2070s at home and when driveing a 4k timeline it's not far off my machine at the studio with dual 2080Ti's
there is a a substaintial diffrence when exporting a 4k master tho
any more VRAM than is needed is a total waste of resource, more is not better, more wastes power, adds heat and gives you zero in return... so really defineing where the line between "not enought" and "too much" is very dependant on workflow and delvierables
i find 8g enough for a UHD timeline with some caches in play
and 11g is enough for aUHD timeline with next to no caches in play
i deliver 4k and UHD so my timelines and monitor match the deliverables
the only wild card is haveing to deliver cam orig, and the cam is Helium
the few times that's been on the table i've sent 8k DNxHR444 back to editorial to test, and their machine melted down... so that became a non issue and i dleivered 4k that they could actualy use to finish
no idea how Pp leverages dual GPU's
i'd get the minimum needed today, and keep my resources avb for when the new tools are reality

Thank you Dermot!

 

[Misha Engel]

Can't say much about PP because we almost don't use it anymore.
For Resolve it depends on the resolution, the effects you're gonna use and if you use CUDA or OpenCL.
With CUDA you need more VRAM because of the different memory handling.
It's pretty easy to crash(running out of memory) a 11 GB 2080ti with 6 or 8k footage running CUDA where a 6 GB RX 5600 XT running OpenCL keeps on going with the same footage without problems.
Here is a quick overview what you can expect from different kinds of GPU's in Resolve 15(same engine as Resolve 16).
https://www.pugetsystems.com/pic_disp.php?id=53796

And as mentioned bevor since Resolve 15 works better with one fast GPU than with multiple slower ones.

 

[Zack Birlew]

I would also say to wait until the 3000 series is out, which should be very soon, and see how they stack up with the current high end RTX products. If you had to buy now, I wouldn't think a single Titan RTX would hold you back at all and leaves room for a second one later if that isn't enough but then you might as well have gone with a Quadro RTX 8000. On that point, I do wonder if dual Quadro RTX 8000 cards with 96gb of VRAM by NVLink would be supported but that's a hefty investment for what could be features in next year's Geforce series.

 

[Michael Hopper]

I agree with waiting a little longer. Current gen is so near EOL it would be tough not to want to kick yourself shortly after-especially if the rumours of up to 50% improvement on top tier consumer cards are true.

Quadro is substantially more money. It would be ideal if you need massive amounts of VRAM, ECC RAM, sync ability for display arrays, and enterprise specific drivers.

From an NLE video editing and rendering performance viewpoint RTX Titan and Quadro 8000 are pretty much neck and neck (same gpu, # of cuda cores, and boost clock speeds, but the Quadro has a slightly higher base clock and twice the memory capacity).

For what it's worth I gifted away my two 2080 Ti's and switched to an RTX Titan (partially because I did have some instances I needed more VRAM). Benefits to single card include much lower system power draw and temperatures across the board, plus my computer lost a noticeable amount of weight and clutter- which is a big deal if you move your workstation frequently.

Also worth noting some of the AIB 2080ti's can outperform the RTX Titan out of the box just with their cooling improvements.

I'm sure you've experienced issues running an SLI setup. NVLink doesn't solve those problems. Depending on what software you use, you might be surprised how seldom you see each card used to its full potential. Budgets vary and maybe it's more feasible to acquire the lesser GPU then buy a second one a little further down the road.

Whichever way you decide to go you will definitely experience a huge bump compared to your current setup. Just be prepared to unearth the next bottleneck in your system (CPU)

 

[Mark K.]

I have a pair of Radeon VIIs in my current build and I'm finding the computer is tapping into most of that 16GB of VRAM all the time. I'd be wary of even 11GB cards these days (since recording resolutions keep on rising). 16GB would be the minimum I'd be looking at.

If that means getting a single beefy card (instead of two smaller ones), I'd go with the beefier card.