Afterburner installed & working but no fps playback benefit?

Hey folks,

I don't know if this is Resolve 18 related or Resolve in general.

tl;dr — I installed an Afterburner and even though it's definitely working, I don't see a difference.

I have a Mac Pro and just picked up a used Afterburner card for half price. Sweet! I work almost exclusively in ProRes, multicam 4K30, and honestly in FCP never had an issue so never bothered getting the card. Resolve though drops a lot more frames and in heavy multicam I often have to let it stutter or close the multicam window, which isn't what I bought this crazy Mac for. So when I found a used card I decided to get it.

I'm running Resolve 18 of course and built a test timeline before I installed the card. My methodology (possibly flawed; you tell me) was to build a timeline wtih 1, 2, 4, 6 then 9 streams of 4K30 (no effects, just the scaled clips on the timeline), then kept pushing it to 12 and finally 16 tracks. I turned off all caches and enabled "show all video frames", hit play and watched the fps counter.

Pre installation of the card, it didn't drop below 29.97 until I hit 9 streams, and that dropped to 27fps. Then with 12 streams it dropped to 19 fps, and at 16 streams, 15.5 fps.

Each stream is about 600Mbit/s, and my MNVMe RAID tests with read speed of 2250 MB/s — so at the max of 16 streams, that's 16 × 600 = 9,600 mb/s which is 1200 MB/s — so just a little over half the bandwidth. This should not be a drive speed issue.

Playing back all 16 streams, iStat Menus reports that the CPU is pegged at 78% and the GPU's CPU is 97%, and GPU's VRAM is at 85%. OK great! I'm maxing out the computer, so an Afterburner should make a real difference, right?

Well, card installed… preference in Resolve set to use Afterburner… the Activity Monitor's Afterburner window shows it working… and playing back that same 16 stream test, oh YEAH you can see the Afterburner cooking (only using about half its power), and the CPU usage on the Mac is a paltry 11%! Amazing.

afterburner installed

CleanShot 2022-05-09 at 16.42.38@2x.png (120.83 KiB) Viewed 1300 times

But? But I'm still getting identical fps playback performance. 27fps with 9 streams, 19fps with 12, and 15.5fps with 16.

Is this what is expected? I literally just installed this so haven't done any real work, and maybe I'll feel that things are faster / smoother as I edit and add fx and grade and so-on. But I'm very surprised that this simple test gave no different results, even though the power load is clearly being distributed. To be clear, I'm never editing with that many streams — that was a stress test. But I do edit five to eight stream mutlicams with color, fusion, etc on the timeline.

So, I guess the questions are… 
1. Am I testing the wrong thing?
2. Since the GPU is still pegged, is that the weak point?
3. If the GPU is the weak point, will I see a benefit from the Afterburner elsewhere in Resolve?
4. What is the meaning of life, anyway?

Thanks guys,
-Joseph

Is the footage actually 30fps or 29.97fps?

You seem to indicate that the timeline is running at 29.97fps but said that the footage is 4K30 - if it is by chance 30fps then Resolve would need to retime the footage to match the timeline for playback; that is the one thing that I can currently think of that is standing out from this description as you seem to have covered the other basics that would come to mind to check.

Have you tried generating proxies to see if that has any effect on the performance?

I'm running Resolve 18 of course and built a test timeline before I installed the card. My methodology (possibly flawed; you tell me) was to build a timeline wtih 1, 2, 4, 6 then 9 streams of 4K30 (no effects, just the scaled clips on the timeline), then kept pushing it to 12 and finally 16 tracks. I turned off all caches and enabled "show all video frames", hit play and watched the fps counter.

If you are merging these clips and showing them at the same time then I think you are likely reaching another bottleneck and are inadvertently testing the wrong thing. Merge operations in Resolve as far as I can tell are single threaded. That would be why the CPU doesn't show much usage despite being a bottleneck.

I always felt the Afterburner card was a strange thing. Whoever heard of anyone not being able to decode ProRes in realtime on a modern workstation?

Is the footage actually 30fps or 29.97fps?

You seem to indicate that the timeline is running at 29.97fps but said that the footage is 4K30 - if it is by chance 30fps then Resolve would need to retime the footage to match the timeline for playback; that is the one thing that I can currently think of that is standing out from this description as you seem to have covered the other basics that would come to mind to check.

Have you tried generating proxies to see if that has any effect on the performance?

It's all 29.97.

I'm running Resolve 18 of course and built a test timeline before I installed the card. My methodology (possibly flawed; you tell me) was to build a timeline wtih 1, 2, 4, 6 then 9 streams of 4K30 (no effects, just the scaled clips on the timeline), then kept pushing it to 12 and finally 16 tracks. I turned off all caches and enabled "show all video frames", hit play and watched the fps counter.

If you are merging these clips and showing them at the same time then I think you are likely reaching another bottleneck and are inadvertently testing the wrong thing. Merge operations in Resolve as far as I can tell are single threaded. That would be why the CPU doesn't show much usage despite being a bottleneck.

I always felt the Afterburner card was a strange thing. Whoever heard of anyone not being able to decode ProRes in realtime on a modern workstation?

No merging. These clips are just stacked one on top of another, getting progressively smaller to fit more streams on screen at once. I agree that the concept of the card seemed odd; when I bought my Mac Pro I didn't get the card because I figured it shouldn't need it. And it was the right choice when I was using FCP. But as I moved to Resolve, I hit more and more situations where playback was choppy with just a few streams and some light color grading or other effects. Since I found a used card for half price I figured it was worth a try. I haven't done much editing in the last week so it's hard for me to say yet whether in real-world use it's making a difference, but for the limited work I've done it does seem to be helping. Could be placebo though; I obviously want this to be worthwhile!

Stacking them to fit them all on screen will be a merge operation, so I think that's it.

Stacking them to fit them all on screen will be a merge operation, so I think that's it.

Oh really? Hm then what’s an appropriate test? I figured if it’s processing a dozen streams of ProRes then it’s processing a dozen streams of ProRes, regardless of their position on screen. And remember in my test I do actually see the card taking the brunt of the work and freeing up the CPU, but I’m still seeing the same FPS playback as when the card was not installed (hence my wondering if that’s the GPU limitation).

-Joseph

I guess my point is in order for this to work better in Resolve, Resolve has to change the way it works.

As you say, the Afterburner is doing what it's supposed to, it's decoding all the ProRes streams and making them available to Resolve. It's after that, the processing in Resolve that is holding back the performance and it has nothing to do with ProRes at that point, it's already decoded into uncompressed RGBA. Resolve has never been particularly efficient at combining video from several tracks.

I guess my point is in order for this to work better in Resolve, Resolve has to change the way it works.

As you say, the Afterburner is doing what it's supposed to, it's decoding all the ProRes streams and making them available to Resolve. It's after that, the processing in Resolve that is holding back the performance and it has nothing to do with ProRes at that point, it's already decoded into uncompressed RGBA. Resolve has never been particularly efficient at combining video from several tracks.

Interesting. And would the same "limitation" affect multicam playback (just playing the multicam window, where you're seeing all angles at once)?

Of that I'm not sure. The only thing I've tested extensively is having clips or generated media on multiple tracks and verifying that the CPU load characteristics points to it being a single-threaded operation. Meaning it doesn't scale well as you add video tracks, even if there are multiple CPU cores idling. It might be more involved than that, RAM copy speed might be a factor too but it's unclear which of those bottlenecks are encountered first.

Would be great if someone from BMD could chime in.

OK, thanks. At the end of the day… I just want this to play smoothly! *sigh*

It would also be interesting to know if your projects would play better on an M1 Max or similar, because of the unified memory.

It would also be interesting to know if your projects would play better on an M1 Max or similar, because of the unified memory.

yeah… I don't intend to upgrade any time soon. Unless Apple offers some great trade-in / upgrade program from the Intel Mac Pro to an M1 Mac Pro whenever that comes out, that won't be happening soon. My Mac Pro has all the PCIE card slots full and so a Studio Mac is of no interest to me! Not to mention that this was no small investment

If your GPU is at the limit, there’s your bottleneck.
AFAIK, the Afterburner is only helping with decoding or encoding ProRes.

Is it a core limitation? Render the media out to 8K and see what happens.

Good Luck

If your GPU is at the limit, there’s your bottleneck.
AFAIK, the Afterburner is only helping with decoding or encoding ProRes.

I think you're right.

Is it a core limitation? Render the media out to 8K and see what happens.

Good Luck

Not sure what you mean by "core limitation"? But if I render to 8K then sure that'll reduce it to a single stream… the point here is to see just what benefit the Afterburner is providing. I think the conclusion is that ultimately it's freeing the CPU for other processes, which I wasn't really doing by just playing back multiple streams of ProRes (and not even hitting the max of the Afterburner card). As the previously commented stated (and I had guessed), the GPU is the limitation in this specific test.

Again all I'm trying to do is figure out if this silly card is worth it

I meant render the 4K footage out to 8K and test it again with 8K media.

The Afterburner card shouldn't be giving the CPU the day off, to do what exactly?

I'm also on an Intel machineI and ran a similar test with a bunch of HD Prores and I noticed that the FPS counter isn't accurate. With 8 layers, playback was a solid 23.976 with performance mode set to auto. With performance mode disabled 23.976 was still reported though the footage was clearly playing back at 6-8fps. I guess I'm saying that make sure what Resolve is reporting is correct, it's not on my end.

Good Luck

I meant render the 4K footage out to 8K and test it again with 8K media.

The Afterburner card shouldn't be giving the CPU the day off, to do what exactly?

I'm also on an Intel machineI and ran a similar test with a bunch of HD Prores and I noticed that the FPS counter isn't accurate. With 8 layers, playback was a solid 23.976 with performance mode set to auto. With performance mode disabled 23.976 was still reported though the footage was clearly playing back at 6-8fps. I guess I'm saying that make sure what Resolve is reporting is correct, it's not on my end.

Good Luck

oh I see, just to push it harder. The Afterburner is taking processing of ProRes off of the CPU so the CPU can do whatever else Resolve needs of it — which in itself is a totally legit question. What does Resolve use CPU for, and what does it use GPU for?

For the FPS counter, what I did was enable "show all frames" and that way the framerate counter drops to whatever the system can handle. Otherwise it'll always show full framerate and just skip as many frames as it needs. At least that's my understanding of it.