Not in a rush. I was fine with the 8 FPS, just wanted to understand more. Perhaps it’s just software optimization in the end. Thanks for the suggestions.
It’s always been software optimization. Taking advantage of multiple cpu/gpu cores is not that simple. I hope they squeeze out everything AS has to offer.
Hmm., with NVidia scaling seems to be well with multiple cores - but then the whole software seems to be best optimised for NV at the moment.
P.S: They had a big speed gain on Apple silicone with the 3.4.0.0.a for SD / generally 2x upscales with Iris - but that’s also diminished by following releases again.
Since there also is a big bug in Iris with Sonoma atm maybe things get better in the future - we‘ll see.
Changing RAM to 10% / minimum can make a huge difference with many GPU Cores - so you’ll have to try that again on the Mac Studio
Thats because Nvidia did invest since 2016 into AI and is the marketleader.
Their GPUs are build all around AI.
And, of course, because they currently have the highest raw processing power.
The above statement wasn’t meant as an accusation / complaint, just reporting the current state (as I believe it is).
Mac Studio arrived today with the 12 CPU core 30 GPU core M2 Max. 100% memory 704x576p to 1440x1080p with the same settings as before = 13 FPS. Limited memory to 10% GPU activity peaked and 22 FPS.
Happy overall with the improvement over M1 8 FPS. Now, if somebody could explain why limiting the memory has such a counter effect?
1 Like
The reply from Topaz in my topic below gives some explanation. I’m glad to hear the 10% trick works so well on your Mac Studio M2.
Thanks.
Andy
2 Likes
Have you tried running a comparison with low power option, it mentions that GPU not used and all the workload is done my Neural Engine and CPU. I have M1 Max, and noticed it little faster actually. Topaz V5.2.3.
This means will it be actually faster with 10% or 20%? What about the low power option?
How do you Check GPU History? I also have M1 Max.
Where is the mac studio?
To view GPU history, open Activity Monitor and use ⌘4.
Depending on the combination of AI model and video resolution, it could be that low power mode is faster than normal power mode. Low power mode will minimise use of the GPU in favour of the Neural Engine.
Thanks.
Andy
I think it was the other way round: setting RAM to 10% forces (forced? Didn’t test all those 5.x and 6.x builds for that) the use of GPU instead of the Neural engine leading to a performance plus on Apple Silicon rigs with many GPU cores.
I think there is some misunderstanding. It’s correct that lowering / minimising the max memory % can “force” the GPU to be used in some circumstances, particularly when upscaling standard definition with Artemis or Proteus. And in these cases the GPU cores are faster than the Neural Engine alone. I should add that it’s not a case of either/or. I discovered, using ASITOP, that the Neural Engine is always in use to some degree in TVAI.
However, lowering the max memory % is completely separate from Low Power mode. Although there are two ways of setting Low Power mode, either only within TVAI or via MacOS (affecting everything, not only TVAI) the effect is the same*. The GPU cores won’t be used unless absolutely necessary because they are much less power efficient. I’ve noticed that when upscaling SD in Low Power mode, the Iris model will still use a small amount of GPU.
I agree that the fact that Low Power mode is faster in particular circumstances is very counter-intuitive. I discovered purely by accident that on the M4 Pro, upscaling SD using Rhea, with max memory set to 100%, is faster using the Neural Engine alone than the GPU / Neural Engine (partial) combination. And the only way to “force” this is to use Low Power mode.
So I think the only way to get the maximum possible performance on Apple Silicon is to experiment with different combinations of settings. Certainly don’t rely any default settings - or intuition!
*- But if MacOS Lower Power mode is activated after the rendering has started, I’ve found that the Neural Engine / GPU combination isn’t (or can’t be) changed. The only thing left for the OS to do in this case is to lower the frequency significantly to reduce power - and fan noise. At least this is the case on the Mac Mini M4 (Pro).
.
Thanks.
Andy
2 Likes
I found that using 10%-20% Memory in M1 Max with 64gb has better results. With Low Power at the Same time. Quality wise.
Photo, not video, processing experience with Lightroom Classic and Topaz AI on a MacBook Pro Max 16 with 96GB of RAM and 4TB SSD. Compared to a previous HP Z440 Zeon system with 64GB of internal storage ($5700 system).
Migrating from the almost 10 year old HP to what was a new MacBook about 18 months ago, the MacBook is mostly faster for things that don’t matter such as booting up (I only do that every couple of days.
Other than AI processes such as the new DeNoise in Lightroom, nothing else is faster and some minor tools are slightly slower. Can’t do an AI comparison because the video card in the HP didn’t support the AI functions in Lightroom or Topaz.
For AI processing on a video card, Macs are at the low end. Depending on the benchmark, the M4 Max is 20-100 times slower than the new Nvidia 5090. And of course we are stuck with our Macs not being upgradable.
The “fantastic” battery life of the Magic (mythical) processors only applies to a light workload. Yesterday I spent 70 minutes in Lightroom and Topaz and the battery level dropped from 80 to 44 percent remaining. This would give 4 hours on a full charge. This is what I’ve seen on the serious video editing forums.
1 Like
