I am curious to understand whether Topaz VideoAI relies on the Neural Engine or GPU. I’ve got M1 Mac minis and an M1 Pro MacBook Pro 16". Apple’s M2 generation improves Neural Engine performance by 40%, and increases the GPU core counts up to 38 cores in the M2 Max CPU (I guess M2 Ultra may come with up to 74 cores Real Soon Now™).
I’m in the market for a new Mac. If VideoAI relies primarily on the Neural Engine, which is the same across M2/M2 Pro/M2 Max, then I’m not going to get any additional performance improvement from spending more money on the more capable chips. A base MacBook or Mac mini would do the trick. But if VideoAI is GPU-driven, then it makes sense to go for the 38-core M2 Max or even to wait for M2 Ultra and its 74 cores.
Any insights to offer? Can anyone offer any comparisons in FPS improvements between different versions of Apple Silicon?
Thanks, Stephen. What were the TVAI Process settings? On my Mac Studio M1 Max I get better performance by specifying the AI Processor instead of Auto together with lowering the memory % to around 25%. I’d be interested if you find something similar. Also, if running 2 processes in parallel, I get the best overall performance running 1 process with memory set to 100% (actually anything above 33%) and the other set to 25% (actually anything below 34%). Counter-intuitive I know, but that’s what I’ve found on my Mac Studio. The actual percentage “threshold” seems to relate to the video dimensions. Above or below seems to direct processing to CPU or GPU respectively.
I specify the M1 Pro/M2 Pro (not auto) in both and set memory to max.
I’ve never tried changing the memory before, I’ll give it a shot.
Not sure which settings you’re looking for, but no noise or grain added. 2 sources, 1 at 480/30 one at 576/25, both scaled vertically to 1080p (4:3), robust dehalo (the one without frame doubling).
Finished my new (win) desktop few weeks ago,13900k + 3090FE + 2tb 4.0 M2 SSD + 16gb RAM…results are better,but i´ve seen the comparsion m2 vs Win in the last beta mailing… benchs with m2 macbook air vs new probook would be interesting
It seems to use the neural engine for the models and the GPU for encoding, from what I can tell.
As for using the GPU “efficiently” I’m not sure the task of ML can be split up, as the options are limited to using the CPU or the specialized feature of the GPU (tensor, neural engine etc) if available, and defaulting to CPU if its not.
1080p source, Enhancement using Artemis Denoise/Sharpen, Low Quality, Noisy input video condition
1.9 fps (sometimes peaks at 2.9 fps)
MacBook Pro M1 Max 64GB
1080p source, Enhancement using Artemis Denoise/Sharpen, Low Quality, Noisy input video condition
7.7 fps (sometimes peaks at 9.1 fps)
Mac Studio M1 Ultra 128GB
1080p source, Enhancement using Artemis Denoise/Sharpen, Low Quality, Noisy input video condition
14.8 fps (sometimes peaks at 17.2 fps)
Edit: GPU Activity is pegged at 94% to 100% while running on each of the machines, so it makes HEAVY and VERY GOOD use of the GPU. These measurements gathered using multiple tools, but my favourite being “GPU Monitor Pro” from the App Store. GPU Monitor Pro on the Mac App Store
Thanks @brock.gunter-smith. Were these tests run recently with the last eat build of Topaz? I’d love to see results in these machines doing upscales from FHD to 4k.
@shao.zhang here are the results of my tests with the same video file that was 1920x1080 at 29.978 fps (34677 total frames) that was upscaled to 4K 3840x2160 using identical settings on each of my computers using Topaz Video AI 3.1.2.
M1 16GB Mac Mini, FHD to 4K upscaling with Proteus model (auto)
Time to convert: 16 minutes 56 seconds (1016 seconds)
FPS while converting: 2209 frames / 1016 seconds / = 2.1742125984 fps
M1 Max 64GB MacBook Pro, FHD to 4K upscaling with Proteus model (auto)
Time to convert: 8 minutes 0 seconds (480 seconds)
FPS while converting: 2209 frames / 480 seconds = 4.6020833333 fps
M1 Ultra 128GB Mac Studio, FHD to 4K upscaling with Proteus model (auto)
Time to convert: 5 minutes 15 seconds (315 seconds)
FPS while converting: 2209 frames / 315 seconds = 7.0126984127 fps
Absolutely. The GPU cores really are where so much work gets done these days in many surprising ways from most apps. The more, the better that’s for sure.
