“Macbook Pro with M4 Max enables: …up to 1.6x faster when compared to the 16-inch MacBook Pro with M1 Max.”
Assuming comparison was with the highest specs - 32 core M1 Max and 40 core M4 Max, I think 1.6x (at most) is really disappointing.
Can Topaz elaborate further on the comparison? For example, do they know if this was mainly using the GPU cores?
I’ve read that the neural engine on the M4 should be much faster than the M1 at 38 vs 11 trillion operations per second, so it would be really interesting to find out for sure.
It’s been awhile since I’ve took the time to compare benchmarks.
Looking only at the max variants and the model I use most: The M2 and M3 Max look to be the same for speed at M2 8.65 fps and M3 8.33 fps (Proteus auto 2X). The M1 does appear to be a little bit slower at 5.08 fps.
The chart someone shared yesterday form the Apple website makes it look like the M4 will be slightly better than the M3. So maybe it’ll get to 9 fps?
According to Activity monitor, on my m1pro Topaz does not use the GPU at all or very much of the CPU. It uses mostly the neural engine. Compared to the m1 version of the neural engine, the m4 version is about 4.7 times faster on Geekbench AI quantized score.
Someone recently uploaded an M4pro to Geek bench AI.
Evidently Topaz uses Ai for its enhancement … so since the Ai has been increased in the M4 compared to fairly similiar Ai in the M1, M2 and M3, then logically video enhancement should be quicker - maybe a lot quicker, on an M4. Perhaps an M4 Pro is quicker than an M3 Max?
Apple M4 max (40) (and ALL other Mx) would benefit from full METAL coding, gpu and neural maxed would have similar output to the nvidia 4080… please please optimize the software.
I mean this is terrible for all Apple M owners… Lol the old M1 is 1/3rd slower ???
This is the uber laptop ATM… and barebones below (M1)
Topaz Video AI v5.4.0
System Information
OS: Mac v15.01
CPU: Apple M4 Max 128 GB
GPU: Apple M4 Max 96 GB
Processing Settings
device: 0 vram: 1 instances: 1
Input Resolution: 1920x1080
Benchmark Results
Artemis 1X: 14.11 fps 2X: 08.11 fps 4X: 02.74 fps
Iris 1X: 13.43 fps 2X: 05.22 fps 4X: 01.87 fps
Proteus 1X: 14.71 fps 2X: 09.31 fps 4X: 02.67 fps
Gaia 1X: 04.18 fps 2X: 02.99 fps 4X: 02.16 fps
Nyx 1X: 05.36 fps 2X: 05.70 fps
Nyx Fast 1X: 11.56 fps
Rhea 4X: 01.10 fps
ORIGINAL M1 MacBook Pro BareBones… This is brutal.
I moved from M1 to M4. I’d love to be able to compare but it’s crashing all the time with the M4.
It’s just too irritating now, I wish there was alternative software.
I decided to get a Mac mini M4 Pro. Fan noise is annoying when the GPU cores are in heavy use. If you want silence, you can choose the Low Power mode in MacOS or within TVAI but you loose at least 40% performance in either case. Or use noise-cancelling earphones.
When using only the Neural Engine, I’ve found the fan is always inaudible regardless of the workload. So that’s a positive.
I don’t think it has anything to do with the amount of money they made from people on PCs. (I love how that implies that Mac is not a personal computer…)
It is Apple’s fault. Topaz has done everything by Apple’s instruction. They have used every tool available to convert their AI models to run as fast as possible on Apple silicon. Whatever speed TVAI gets, is what Apple has allowed.
I’m under the impression that Topaz doesn’t use Metal and doesn’t take advantage of the full potential of Apple Silicon.
Both may be true I suppose; that it is Apple’s fault that Topaz hasn’t been rewritten for Apple Silicon.
However, it’s true that TVAI doesn’t use the GPU at all. I’m stunned by what happens when it does …
<i know the example below isn’t apples to oranges for many many reasons but nevertheless …>
I’d encourage anyone who has a mac to go ahead and download Unsqueeze from the App Store. Upscaling to 4k is completely free. My jaw dropped when my four year old M1 Max upscaled a 30 second long 1080p video in less than 3 seconds. Now it’s not using different parameters, it’s not downloading different models but … 3 seconds?
Anyway. I do believe that were TL to rewrite taking advantage of Metal and Apple Silicon we would see 10x speeds.
Topaz has been training their AI models since before Apple made their version of AI available. Let’s pretend that Topaz was using PyTorch for this. Since they have been doing it for so long, there would be significant/insurmountable cost to start fresh on another framework. Apple, knowing this would be the case, provided tools to convert AI models from popular frameworks like PyTorch to theirs.
There are several possible reasons why that conversion might not lead to full use of the hardware. The documentation for Core ML Talks heavily about things like natural language. Over all, I got the impression that they focused mainly on large language models with their AI engine. Though there is a section on ‘Vision’ it is possible that videos aren’t it’s strong point. If this is the case, then starting fresh, using only Core ML, would give the same speed and usage results.
It would not be against what Apple has done in the past, to intentionally give degraded performance for outside-Apple-origin programs. (This is pure conspiracy from me.)
Or it could be that the tool just has limits on how well it can convert models. If either of those last two are true, then starting fresh with Core ML would yield awesome amounts of speed.
Apple Metal is their version of Nvidia CUDA. TVAI runs faster on Nvidia TensorFlow than on CUDA. (Assuming that’s what Topaz is using with their claims of using the Tensor Cores in Nvidia cards. [That would make it logical to use Apple Core ML rather than Apple Metal.])
I know this is a Topaz Video AI thread, but it’s truly depressing that Super Focus on my M1 Pro is estimating 58 minutes of processing…that’s after the almost 5 minute time just for the estimate. I see lots of finger-pointing in this thread, and I don’t know if it’s Topaz Labs’ fault or Apple’s fault, but the end result is Topaz Labs’ software is the only software that truly makes my laptop feel slow.
Fingers crossed that by the time M5 laptops and desktops come out, some speed gains will have been realized that will make Super Focus actually usable.
No. As seen before they actually had done some Mac optimization in the past with the Iris model just flying afterwards (being faster than Proteus) - and then reversed this due to some strange artifacts in a new MacOS version.
After that rough bug fix the Iris 2x runs only about half the speed of what would be possible and this is still the same even a year later (even though the old model now runs without artifacts again if you still have it.
This is just one example.
Also the fact that TVAI quite often runs faster on Apple Silicon with RAM set to minimum or sometimes even in low power mode (!!) shows that there is about no optimisation for speed at all, likely not even testing.
That’s not how I remember it. They just knee-capped the speed because the new version of Mac OS was making those artifacts. It was never noticeably faster than any of the other Mac optimized models.
True that they left the knee-cap limitation in place for no reason. It might even still be there.
Tried it. (Mac Studio M2 Max). Yes it is super fast. But the result is terrible. No, absolutly no denoising or extra sharpening. No idea for what this app should be usefull.
and barebones below (M1)
