Davinci Resolve 10.1.3 performance report on DRAGON footage with 8C nMP

[Michael Tiemann]

In case anybody is curious, I benchmarked my new Mac Pro system (8C dual D700 64GB RAM TB RAID array) running the latest Davinci Resolve against some Dragon footage. First and foremost, Resolve has a problem with the latest (5.5.47 drivers) so the older (5.5.28) drivers are needed. That said, with 6K FF footage and three color correction nodes, I could get real-time playback by downsizing the project resolution to 4K DCI, Half Res Good decode and Quarter Res Good playback. The decoder process is definitely the rate-limiting step: when I changed the project resolution to 2K DCI and set decode and playback to Half Res Good, performance was no better than when the project size was 4K DCI.

On the other hand, when the project size was set to 6K FF (6144x3160), I could only get real-time playback with Quarter Res Good playback and zero nodes. The first node took the playback speed down to 22fps, the second node down to 18 fps, and the third node down to 14 fps. In the three node case the CPU utilization doesn't go much past 650%, so not CPU limited but GPU limited.

Anyway, just some data for people wondering what they can expect from a nMP without RR-X.

 

[mikeburton]

Hey Michael,

Thanks for reporting. If it's not too much trouble would you mind setting up a 1080p project @ 23,98fps and rendering Dragon at Full Res, half res high and half res good to ProResHQ and report back? Would be quite helpful.

 

[Michael Tiemann]

For full-res premium decode to 1080p 422 ProRes HQ with 2 channels of audio, 30 seconds of content takes:

1:34.93 at full-res premium
1:34.82 at half-res premium
1:09.70 at half-res good
0:27.79 at quarter-res good

CPU saturated at 1600% utilization in most cases, but it looked like it might have been falling off to 1200% in quarter-res good. All of the above is with three nodes active.

 

[Iskandar Abdullayev]

I think any Dragon camera purchase should include Rocket-X - RED is giving a hefty discount on it

 

[Nick Timmons]

Unless you're a DIT pumping out transcodes, GPU performance in Resolve or RCX or whatever makes a pretty compelling argument against the Rocket-X, and doesn't cost nearly as much.

 

[Nick Pasquariello]

For full-res premium decode to 1080p 422 ProRes HQ with 2 channels of audio, 30 seconds of content takes:

1:34.93 at full-res premium
1:34.82 at half-res premium
1:09.70 at half-res good
0:27.79 at quarter-res good

CPU saturated at 1600% utilization in most cases, but it looked like it might have been falling off to 1200% in quarter-res good. All of the above is with three nodes active.

Michael, can you state what those three nodes are each adjusting? Noise reduction taxes the system differently than a contrast curve, for example. (Either way, the numbers are much appreciated, so thanks for taking the initiative to test.)

 

[Michael Tiemann]

Michael, can you state what those three nodes are each adjusting? Noise reduction taxes the system differently than a contrast curve, for example. (Either way, the numbers are much appreciated, so thanks for taking the initiative to test.)

They are all very simple primary corrections with qualifiers. No noise reduction / object tracking / power windows or anything of the sort. Just lift / gamma / gain and saturation based on HSL in the image.

 

[Tommaso Alvisi]

Michael, thanks for the test!!