18.6 vs 18.5 Studio export tests/results AI optimization

18.6 vs 18.5 Studio export tests/results AI optimization

Results:
Speed Warp - 2x faster (15 fps vs 7 fps)
Noise Reduction - no change!
Depth Map - 30% faster

Would be great if BMD could speed up Noise Reduction with the Neural Engine.

Fantastic optimization in Speed Warp.
Thanks once again to BMD for the continued upgrades to Resolve.

Feel free to post your own results/data.

Process:
Resolve 18.5 Studio.
Updated Nvidia Studio driver to current 537.42
Tested 18.5.
Installed 18.6
Resolve ran 'Optimize Davinci Neural Engine' 7 min

source clip: Mavic 2 Pro / UHD 3840x2160 / 29.97 fps / 100 mbps / 4.2.0
30 second clip
Computer specs in signature
Export Settings:

0 export settings.jpg (69.7 KiB) Viewed 1451 times

Data

18.5:

1) Speed Warp 50% slow (= 1:00 clip)
+ 2160p export : 4:13 = 7 fps
+ 1080p export : 0:58 = 31 fps

2) Noise Reduction / TNR, 5 frames, better, 10 / SNR, faster, medium, 5
+ 2160p : 0:56 = 16 fps
+ 1080p : 0:14 = 64 fps

3) Depth Map (*interesting note: resolution didn't affect fps!?)
+ 2160p : 0:46 = 19.5 fps
+ 1080p : 0:46 = 19.5 fps

18.6:

1) Speed Warp 50% slow (= 1:00 clip)
+ 2160p export : 2:00 = 15 fps
+ 1080p export : 0:33 = 55 fps

2) Noise Reduction / TNR, 5 frames, better, 10 / SNR, faster, medium, 5
+ 2160p : 0:56 = 16 fps
+ 1080p : 0:14 = 64 fps
** no change **

3) Depth Map (*interesting note: resolution didn't affect fps!)
+ 2160p : 0:43 = 21 fps
+ 1080p : 0:42 = 21 fps

Thanks for the test.

Glad people are seeing real results, not just marketing claims.

Was the AI upscaling supposed to be improved too?

Agreed NR is the biggest pain point. I do hope BM is working on a ML version of it to compete with Neat and Topaz.

Adobe added their new NR to Lightroom a few months ago and it is another world vs their old tool on stills. Slow though.

I get 3 fps with Neat on my machine, so it is always a question of how bad do you want it

3) Depth Map (*interesting note: resolution didn't affect fps!)
+ 2160p : 0:43 = 21 fps
+ 1080p : 0:42 = 21 fps

Detph map most probably downscales the input to fixed resolution, either for performance reasons on because model expects fixed-size input tensor. Result is later upscaled to original res again.

...Was the AI upscaling supposed to be improved too?...

Puget did a great suite of testing and shows AI upscaling as 60% faster in 18.6

Worth a read and some great charts:

https://www.pugetsystems.com/labs/artic ... mizations/