M3Ultra Studio vs M1Max Davinci Resolve Voice Training Speed = Little Difference

[smithdr]

Hi All:

I thought that I would start a thread for those considering an M3 Ultra Studio to perform AI Voice Training in Davinci Resolve Studio. I have both an M1 Max MBP 16 10 Core CPU / 32 Core GPU / 64 GB RAM and an M3 Ultra Studio 32 Core CPU/60 Core GPU/256 GB RAM. I performed Davinci Resolve Voice Training a on 32 Bit / 48,000 Hz / 5 minute wave file using the Better Option. I used iStat Menu to determine what resources were being used.

In general, the M3 Ultra is 80 - 90% faster than the M1 Max on tasks such as encoding video for production and face tracking using the face refinement application. However, the Voice Model Training M3 Ultra shows little performance gains over the M1 Max. Here are the results:

The M3 Ultra shows CPU utilization to be about 85% with all 28 cores performing some task. The GPU also shows about 86% utilization. There is no way using iStat Menus to determine how many cores are being utilized. Voice training was completed in 1 hr 5 minutes.

The M1Max shows 65% CPU with all cores being utilized and about 55% GPU utilization. The M1Max completed voice training in 1 hr 12 minutes.

I am really surprised by this result and double checked my work to make sure I did not make an error. The M3 Ultra showing far higher core utilization with it’s more cores while the M1 Max shows lower utilization with lower cores only results in a 11% improvement in training speed for the M3 Ultra.

I post this for commentary and to help others considering purchase of an M3 Ultra for use with AI Voice Training in Davinci Resolve.

Don

 

[haralds]

I will be sticking with my at the time pricy M1 Ultra Studio for a few years to come. It's a workhorse!

 

[smithdr]

I will be sticking with my at the time pricy M1 Ultra Studio for a few years to come. It's a workhorse!

If you already have an M1 Ultra that is a good solution. I was working on an M1 Max MBP16 and wanted an upgrade. I tried an M4 Max MBP16 and while faster, I found the machine got too hot and was chewing up batteries. So I thought why not get an M3 Ultra, access it from the M1 MBP16 using high performance screen sharing as I prefer working on laptops in my recliner vs being hunched over at a desk--best of both worlds. Generally, this experiment is a success and much faster at the things I repetitively do. But, I am still disappointed and surprised by the M3 Ultra’s Voice Training performance compared to the M1 Max MBP16.

 

[wirefall]

Mac Studio M3 Ultra: What is your recommendation from your experience with Davinci Resolve regarding RAM, 96 or 256 GB? Couldn't find a dedicated comparison so far. Of course, "more is better", but 256GB would be another 2 grand. On the other hand I am willing to pay that if a certain gain in performance in a heavy multitasking setting could justify it. Moreover 256 GB is maybe more future proof. Thanks in advance for your help

 

[smithdr]

Good question. I went with 256 GB of RAM believing this would be necessary with all of the AI applications that I anticipated coming in the near future. As you can see from my experience with Davinci Resolve AI voice training all that additional RAM and GPU (32 cores to 60 cores) had little impact on performance in this AI application in Davinci Resolve.

If I had to do it over again, I would have gotten the 96 GB instead of 256 GB of RAM as I believe it has modest impact on Davinci Resolve for all the things I do including modest Fusion.

 

[wirefall]

Thanks a lot, this was exactly my thinking. AI in Davinci should benefit from more RAM. And what do we know will show us progress in all our apps in 2 or 3 years. There is quite some dynamic regarding AI these days. Interesting enough your results clearly show the facts. So I have to think twice regarding 256 or 96 GB of RAM.

Other opinions?

 

[smithdr]

I also believed that more RAM and CPU cores would significantly improve AI applications in Davinci Resolve. Unfortunately, it had minimal impact. On the Davinci Resolve Forum another user reported a 50% reduction in time running a voice training model using an M4 Mini Pro vs an M1 Max. It appears that Black Magic is using some other functionality associating with the M4 chip that is not present on the M3 and earlier versions.

It is not as though the M3 Ultra is waste of money relative to the M4 Max. When I thought about it, how many voice training models do I need to run? For me, not very often. And once it is done, it is done and I do not need to run it again because the model is preserved.

The M3 Ultra is still 50% faster on things that I do all of time: rendering clips, face refinement, noise reduction, exporting a video, etc. This is because of the extra CPU cores and twice as many video engines. I just did not need 256 GB of RAM to get this improvement--96 GB would have been fine.

Hope this helps.

 

[wirefall]

Yes, indeed very helpful, thanks a lot. You name the core things and it is reasonable to buy a M3 Ultra to have them about 50% faster.

Regarding RAM: Maybe you could check Activity Monitor on your Mac Studio how much RAM is used by a typical workflow with Davinci Resolve. Or maybe iStat Menus in case you have it. Would be nice to know if Davinci could utilize (quite some) more than 96 GB of RAM at all. If so 256 GB could make sense. Also to be future-proof, as such a machine should be good for quite some years.

Thanks again in advance!

 

[smithdr]

I did. On my M1Max I was 48 GB of 64 GB. On the M3 Ultra I am using 48 GB of 256 GB when doing a 5 minute voice training model in best mode.

 

[wirefall]

Interesting.How much RAM consumption do you see doing the obvious suspicious stuff like rendering, noise reduction or exporting?

 

[smithdr]

This would be difficult to do this different clips and Fusion compositions have different resource requirements at would not translate well to other situations.

 

[wirefall]

Of course. My question is more meant just to get a sense of a (more or less) typical amount of RAM your Mac Studio M3 Ultra uses in your daily work. E.g. when it comes to let's say about 60-80 GB (Davinci Resolve etc., not AI / LLM), it wouldn't make sense to buy the 256GB version. On the other hand when you see something like 150GB or more it could make sense. Even more as I think about using LLMs. But of course we talk about another 2 grand to have 256 GB So thanks again for any hints from your experience!

 

[smithdr]

Hi WF:

I exported a 5 minute video from Davinci Resolve Studio (DRS) for YouTube/Facebook. This video had several Fusion compositions and a Magic Mask. During the export, DRS consumed approximately 30 GB of the 256 GB of RAM available.

Hope this helps.

Don

 

[wirefall]

Hi Don,
thanks a lot, much appreciated! So that means 96 GB RAM would be enough.

I remember another posting somewhere (can't find it right now), a user reports more than 150 of 256 GB used by Davinci Resolve and some heavy multitasking around. But that was due to macOS, which "eats up" RAM while running it for some time. So 256 GB were not really being used, there is also some old stuff still remaining in RAM because it just could due to the large amount of RAM. Haven't heard so far about this phenomenon and have to dive in deeper, even if there is some benefit.

However, still being tempted to order 256 GB just to be future proof for quite some years such a machine should last. 🤔🤪
Question is also if there is some sweet spot regarding reselling it some day. Maybe 256 GB is relevant to offer still enough performance in 4-6 years.

 

[Basic75]

I am really surprised by this result and double checked my work to make sure I did not make an error. The M3 Ultra showing far higher core utilization with it’s more cores while the M1 Max shows lower utilization with lower cores only results in a 11% improvement in training speed for the M3 Ultra.

Perhaps the application simply doesn't scale that well. If it spends too much time in critical sections the synchronisation overhead can eat up the additional processing power. Just one theory.

 

[smithdr]

Hi Don,
thanks a lot, much appreciated! So that means 96 GB RAM would be enough.

I remember another posting somewhere (can't find it right now), a user reports more than 150 of 256 GB used by Davinci Resolve and some heavy multitasking around. But that was due to macOS, which "eats up" RAM while running it for some time. So 256 GB were not really being used, there is also some old stuff still remaining in RAM because it just could due to the large amount of RAM. Haven't heard so far about this phenomenon and have to dive in deeper, even if there is some benefit.

However, still being tempted to order 256 GB just to be future proof for quite some years such a machine should last. 🤔🤪
Question is also if there is some sweet spot regarding reselling it some day. Maybe 256 GB is relevant to offer still enough performance in 4-6 years.

For DRS 96 GB is more than enough. There is no such thing as future proofing. Buy the lowest speced computer that you can to get the job done. In this instance, BlackMagic is using something other than RAM to improve the speed on AI operations. Perhaps the neural engines? There is no future proofing for this. It is hardware capability issue.

But if you are hell bent on getting a M3 Ultra with 256 GB RAM I will be happy to sell you mine. 😀

 

[b0fh]

I don't know if you're aware of this, and I don't know if it's even useful...

GitHub - finnvoor/yap: 🗣️ A CLI for on-device speech transcription using Speech.framework on macOS 26

🗣️ A CLI for on-device speech transcription using Speech.framework on macOS 26 - finnvoor/yap

github.com

Apple's New Transcription APIs Blow Past Whisper in Speed Tests

Apple's new speech-to-text transcription APIs in iOS 26 and macOS Tahoe are delivering dramatically faster speeds compared to rival tools, including OpenAI's Whisper, based on beta testing conducted by MacStories' John Voorhees.

www.macrumors.com

 

[smithdr]

I don't know if you're aware of this, and I don't know if it's even useful...

GitHub - finnvoor/yap: 🗣️ A CLI for on-device speech transcription using Speech.framework on macOS 26

🗣️ A CLI for on-device speech transcription using Speech.framework on macOS 26 - finnvoor/yap

github.com

Apple's New Transcription APIs Blow Past Whisper in Speed Tests

Apple's new speech-to-text transcription APIs in iOS 26 and macOS Tahoe are delivering dramatically faster speeds compared to rival tools, including OpenAI's Whisper, based on beta testing conducted by MacStories' John Voorhees.

www.macrumors.com

Thank you. But but not certain how this applies. I am not doing transcription. What I am doing is cloning my voice. Then re-recording the areas where I have made errors and using the clone to repair the errors.

 

[Ethosik]

In general, the M3 Ultra is 80 - 90% faster than the M1 Max on tasks such as encoding video for production and face tracking using the face refinement application. However, the Voice Model Training M3 Ultra shows little performance gains over the M1 Max. Here are the results:

I expect most of this is going from Max to Ultra. I expect you would experience similar gains with the M1 Ultra as well especially for the encoding portion. I have not noticed a difference in that between my M1 Ultra and M3 Ultra test (I returned it and just got another M2 Ultra instead for a little cheaper).