2 x nVidia 980Ti issues

I've got a Asus x99 Pro board with two nVidia 980Tis, Windows 8, Resolve 11.3. With 'Use Display GPU for Compute' off I can hit 25fps with r3d and a few nodes in a 3840x2160 timeline at half debayering good. With it turned on this drops down to around 13-14 fps.

I've also taken one GPU out to test things and it appears I'm not gaining any significant performance having the two cards in there. In fact I can still achieve 25 fps with the one card at half debayer good in the same timeline. The second GPU is not adding much value at all.

So could anyone enlighten me with what might be the issue here? If one card can do it all (display, debayer and compute), surely adding the extra card should make things fly? The biggest reason for me upgrading to the full version of Resolve was so I could tick the 'Use Display GPU for Compute' box to really boost performance. Are the two GPUs not talking to one another correctly?

I have the 980 ti myself but my understanding is that you can't stack cards and expect the Vram, which is what Resolve runs on, to aggregate. i.e. Resolve will only benefit from a single card.

Interesting, although I wouldn't expect performance to be hindered. And if this is true why is the 'Use Display GPU for Compute' a only available in the paid version.

I'd like to know the thoughts of the Resolve technicians on this. On the product page there is plenty of talking up of multiple GPU support including up to 8 GPUs on one system. Is this the Linux system only? Is this able to split the VRAM in someway that I'm not able to with a Windows system?

Right now I'm struggling to see the benefit of the £700 I've paid for the full version as the only difference I'm getting is use of the denoiser. I've got a feature grade on RED lined up in a few weeks which is why I thought the full version would benefit me.

What X99 board and what CPU are you running?

Most x99 boards will drop the speed of the PCI-E slots if you use multiple cards.

For instance, running the X99 Deluxe with a 40 lane CPU will let you run 2 cards at full speed (x16) but if you are running the same board with a 28 lane CPU the 2nd slot will drop to x8 and so the performance from any card in that slow will be impacted.

You could try setting connecting your display to card 2 and then setting Resolve to not use GUI for compute so it's only using the card in the x16 slot. This should give you the full power of that card.

Hi Adam,

I'm using a Asus x99 pro with a 5820k. I'm aware the 5820k only runs at 28 lanes so each card is restricted to 8 x 3.0 PCI E lanes which is plenty enough. As far as I am aware there is no card on the planet that can saturate 8 x PCI 3.0. I certainly wouldn't expect a drop from 25fps to 13fps.

The whole point of me upgrading to the full version was to take advantage of the display GPU for compute, as there should be plenty of juice left over in a 980Ti after taking care of the display. There obviously is as one on its own can do the job by itself. The problem being combining the two.

Both GPUs only show up in system overview when Resolve is started with Use Display GPU for compute turned on. Is this the correct behavior?

Just disabled one GPU in nVidia control panel and I'm getting 25fps Full Res r3d playback which drops after 3 or 4 nodes applied. Re-enabling both GPUs and using display for compute drops playback to 13-14fps and without it Display GPU for compute enabled playback returns to 25fps. However I'm still seeing the same slight drop from after 25fps after applying a few nodes - same as using one GPU. No performance increase. For all I know it is just using one GPU when using Display GPU for Compute is not selected as only one GPU shows in system overview.

Which sockets do you have the cards in? According to the manual when using that CPU the top slot is 16x, then I think PCI-e 16_3 becomes 8x and PCI-e 16_2 becomes 4x(this is a bit unclear as usual with Asus), although the PCI-e 16_2 slot also seems to share with a PCI-e 1x slot and some USB 3 sockets, so it might be worth moving the second GPU into the other slot to see if it makes a difference.

Ah yes! Thank you, Adam. It turns out the x99-pro board does a few weird things. Slot 3 should be a PCI 2.0 x 16 which should be more than plenty but when the board's on -board wi-fi is enabled this gets reduced down to PCI 2.0 x 1. This is not mentioned in the manual but is in the small print on their website. Moving the second GPU to the next slot has sorted things out. Playback is fantastic.

That was a pain in the ass, especially as I wanted the x99 deluxe which wasn't in stock at the time. I believe that board would not have had the same issue.

I've also taken one GPU out to test things and it appears I'm not gaining any significant performance having the two cards in there. In fact I can still achieve 25 fps with the one card at half debayer good in the same timeline. The second GPU is not adding much value at all.

I think you've hit on a fact that people around here don't seem to want to acknowledge which is that just packing your computer with GPU's doesn't necessarily achieve miracles. When I spoke with Blackmagic's tech guys at NAB, they told me that something like, say the Titan X, is so powerful that there is absolutely no advantage to having a second card in there for GUI, and having a second Titan X in there would also be kinda useless unless you have an extra $1k burning a hole in your pocket. While offloading computational tasks to a GPU sounds great on paper, there are limits to how much can be done or else we'd all be buying CPU-less motherboards and just add GPU's as we need them (and can afford them) to handle as little or as much computing as we need.

I think the most computationally-intensive task in Resolve is TNR, so I'm sure a faster card with huge amounts of memory helps there. But for basic color grading nodes and playback, most GPU's can handle quite a lot. For instance, on my considerably modest 2gb GTX 770 I can playback 4K CDNG footage with a handful of color nodes applied at 24fps without any issues whatsoever.

P.S.

The speed of your media drives are really crucial when it comes to good playback. A 4-drive RAID is pretty much a must-have as far as I'm concerned.

P.P.S.

Also, the type of CODEC that you're using has a huge impact on playback as well. For instance, I find 4k CDNG to be silk smooth, but I suspect ProRes4444XQ at 4k might be an issue.

Hi Kays. Thanks for the input. All is good now, with the GPU in the right slot I'm getting some amazing results, much better than with a single GPU. The reason I need the GPUs is for r3d debayering, with the second card in there I can pile on 10+ plus nodes and still hit 25fps with a half res debayer. I've sold in this in to a director as a major factor, being able to grade realtime from the r3d without any need for proxies.

You're quite right about the codec though. I can play the r3d back but I've got some UHD h264 from a drone which Resolve struggles with.

For the uhd just turn the input cache on.

Better to transcode H264 footage to a more edit friendly codec. H264 is fine for online delivery but it's not ideal for frame based editing/grading. Needs to much resources to decide on the fly.

...
You're quite right about the codec though. I can play the r3d back but I've got some UHD h264 from a drone which Resolve struggles with.

Have you checked CPU usage when using this footage as decoding of this type of footage is CPU dependent. It also may be getting decoded using QuickTime which again is a limiting factor on the PC side as it's only 32bit and can slow down decoding even on a very fast system.

Better to transcode H264 footage to a more edit friendly codec. H264 is fine for online delivery but it's not ideal for frame based editing/grading. Needs to much resources to decide on the fly.

If you have the cache on, the resolve will convert the h264 in the preferred cache file (dnx/prores/dpx) and then shirk from those.

Our di it explicitly and convert it.