MP 14,8 M2 Ultra and Davinci Resolve

[Dopemaster]

Was hoping some savvy posters might be able to offer some insight on optimal M2 Ultra configs for working (primarily) in Davinci Resolve.

Use case: feature length projects, 6-8K R3D, ArriRaw, Braw, ProRes etc. Offline/online editing, stabilization, optical flow, neural engine, magic mask, NR, sharpening, upres, colour grading.

Other Apps: Pro Tools, Topaz, After Effects, Nuke, various ML models.

Hard to tell whether opting for more GPU cores or RAM is better ‘bang for you buck’ for video work, given the RAM pool doubles as GPU VRAM.

Future proofing also a consideration. In my view ML tasks is the area that that will become more demanding sooner.

Any insight much appreciated!

 

[elfamosisimoJON]

You can seen M2 Ultra benchmarks for Topaz Video AI in their forums, they have either 60 and 76 cores and 64gb and 128gb ram.

 

[Dopemaster]

Any other takers?

 

[h9826790]

Future proofing also a consideration. In my view ML tasks is the area that that will become more demanding sooner.

I don't have any benchmarks on hand, but my understanding so far. M2 Ultra 192GB is a very decent home use ML machine.

It's GPU isn't that strong (even considered quite weak for model training), however, for the same price, you can't buy anything with that amount of usable VRAM.

It's been tested that M2 Ultra can run a whole ~135GB GPT model without any issue. For personal computer, it's practically impossible to build a computer with 144GB VRAM for that price.

The M2 Ultra isn't fast at all. Especially lack of upgradability, means it will be outperformed by some much cheaper PC parts in a few years. So, really hard to have any future proof. However, anything that need huge amount of VRAM can be benefited by that 192GB unified memory right now.

TBH, if I want a ML machine now, I will simply get the 192GB M2 Ultra Mac (most likely the Mac Studio but not Mac Pro. Mainly because lack of future proof. So, no point to pay that more). Of course, no CUDA, far from ideal. But it really depends on usage. Also, the community recognised this effect, and started to develop software for Apple Silicon ML. By considering the high end gaming GPU still at the 24GB VRAM rage. The M2 Ultra should at least able to run most of the popular ML model in the next few years. It will be slow, but at least "can run". And it will up to the user to decide if they want a faster machine by that time.

 

[Dopemaster]

I don't have any benchmarks on hand, but my understanding so far. M2 Ultra 192GB is a very decent home use ML machine.

It's GPU isn't that strong (even considered quite weak for model training), however, for the same price, you can't buy anything with that amount of usable VRAM.

It's been tested that M2 Ultra can run a whole ~135GB GPT model without any issue. For personal computer, it's practically impossible to build a computer with 144GB VRAM for that price.

The M2 Ultra isn't fast at all. Especially lack of upgradability, means it will be outperformed by some much cheaper PC parts in a few years. So, really hard to have any future proof. However, anything that need huge amount of VRAM can be benefited by that 192GB unified memory right now.

TBH, if I want a ML machine now, I will simply get the 192GB M2 Ultra Mac (most likely the Mac Studio but not Mac Pro. Mainly because lack of future proof. So, no point to pay that more). Of course, no CUDA, far from ideal. But it really depends on usage. Also, the community is recognised this effect, and started to develop software for Apple Silicon ML. By considering the high end gaming GPU still at the 24GB VRAM rage. The M2 Ultra should at least able to run most of the popular ML model in the next few years. It will be slow, but at least "can run". And it will up to the user to decide if they want a faster machine by that time.

Appreciate the response! I found this article enlightening too re ML in Nuke/CopyCat: https://www.fxguide.com/fxfeatured/apple-m2-studio-learns-nuke-machine-learning/

 

[Yebubbleman]

Was hoping some savvy posters might be able to offer some insight on optimal M2 Ultra configs for working (primarily) in Davinci Resolve.

Use case: feature length projects, 6-8K R3D, ArriRaw, Braw, ProRes etc. Offline/online editing, stabilization, optical flow, neural engine, magic mask, NR, sharpening, upres, colour grading.

Other Apps: Pro Tools, Topaz, After Effects, Nuke, various ML models.

Hard to tell whether opting for more GPU cores or RAM is better ‘bang for you buck’ for video work, given the RAM pool doubles as GPU VRAM.

Future proofing also a consideration. In my view ML tasks is the area that that will become more demanding sooner.

Any insight much appreciated!

Disclaimer: I don't use those tools, so my take is speculative at best.

Apple seems to act as though system RAM might as well be discrete VRAM on a traditional discrete GPU. When asked about the GPU in the M2 Ultra, they comment as such. How much that holds in practice, I couldn't say. Certainly, I don't feel like I have 8GB or 16GB of VRAM in my M1 Macs, but that could also be because that the standard M1's GPU is only on par with an NVIDIA GTX 1060.

My guess, based on (a) the differences between 48 and 64 GPU cores in the M1 Ultra, (b) the differences between 30 and 38 cores in the M2 Max, and (c) how the M2 family has done a much better job of scaling up graphics performance with more cores (on higher-tier SoCs), you might be better off upgrading RAM rather than the GPU core count. I might be wrong about that and certainly, I'd implore you to do more research. But, 16 GPU cores isn't that much relative to already having 60 and, like M1 Ultra, software will need to be optimized to fully utilize M2 Ultra (given that it's still two M2 Max SoCs running in parallel with extremely high bandwidth between them).

 

[Dopemaster]

Disclaimer: I don't use those tools, so my take is speculative at best.

Apple seems to act as though system RAM might as well be discrete VRAM on a traditional discrete GPU. When asked about the GPU in the M2 Ultra, they comment as such. How much that holds in practice, I couldn't say. Certainly, I don't feel like I have 8GB or 16GB of VRAM in my M1 Macs, but that could also be because that the standard M1's GPU is only on par with an NVIDIA GTX 1060.

My guess, based on (a) the differences between 48 and 64 GPU cores in the M1 Ultra, (b) the differences between 30 and 38 cores in the M2 Max, and (c) how the M2 family has done a much better job of scaling up graphics performance with more cores (on higher-tier SoCs), you might be better off upgrading RAM rather than the GPU core count. I might be wrong about that and certainly, I'd implore you to do more research. But, 16 GPU cores isn't that much relative to already having 60 and, like M1 Ultra, software will need to be optimized to fully utilize M2 Ultra (given that it's still two M2 Max SoCs running in parallel with extremely high bandwidth between them).

That's the way I'm leaning based off the RW testing that is available out there. Will opt for 128GB Ram, just waiting on more direct comparisons between the GPU options before pulling the trigger.

 

[giffut]

For any video related work the Studio is great, at any option. If you can miximize it, go for it. But as
h9826790 mentionned, it´s always dead end upgrade wise. If you stay within macos based software, you shlould be fine for years.

But: For serious machine learning work you need to get Cuda support and that basically rules out any mac. Basically: built or buy a top end ryzen based epic/threadripper system with nvidia gpus, with tons of ram possible to utilize.

 

[Dopemaster]

For any video related work the Studio is great, at any option. If you can miximize it, go for it. But as
h9826790 mentionned, it´s always dead end upgrade wise. If you stay within macos based software, you shlould be fine for years.

But: For serious machine learning work you need to get Cuda support and that basically rules out any mac. Basically: built or buy a top end ryzen based epic/threadripper system with nvidia gpus, with tons of ram possible to utilize.

Video work is number 1 priority. But ML is increasingly incorporated into video software like Nuke and Resolve, so it’s a secondary consideration. These two softwares are increasingly optimised for Apple Silicon.

Aware that a custom PC would be cheaper and more upgradable, also Nvidia/Cuda currently much better for broad use ML, but not interested in switching back to Windows or Linux purely for that secondary consideration.

 

[chfilm]

I'll get my hands on one next week, as a Resolve editor/colorist. I currently have a 2019 mac pro and can run some tests for you if you tell me what you want to know.

 

[Dopemaster]

I'll get my hands on one next week, as a Resolve editor/colorist. I currently have a 2019 mac pro and can run some tests for you if you tell me what you want to know.

Appreciate the offer! I’m less interested in straight render times, and more interested in timeline performance w/ R3D, ArriRaw, Braw footage. With things like stabilisation, NR, motion blur, magic mask, multi-node colour work etc. Really push it to its limits and see how linearly the gpu scales, also how much Ram Resolve will actually utilise.

Not sure if you’re also a Nuke user but would love to see how it handles Copycat ML in comparison to the 2019 Mac Pro.

 

[PowerMike G5]

As a long time Premiere Pro (FCP 7 and Avid MC before that) user, I decided to randomly test out DaVinci Resolve on my 7,1 the other day. I was surprised by how well optimized the software is in its current state!! I know Premiere has a reputation of not having the greatest performance, but I didn't expect it to be this much compared to DR!

I tested out some Canon Cinema RAW light files, ProRes, 10-bit XF-AVC, HEVC. It all plays so smoothly in DaVinci Resolve. Premiere will literally max out all 16-cores (and the additional 16 hyperthreading cores) on my machine just to playback one stream of 4K DCI Canon CRM. And it still stutters on full res playback.

I load it into Resolve and it barely uses 3 cores and it plays back at full resolution with no problems. Wow!

I'm seeing DR is so heavily GPU-optimized. I'm curious now of the M2 Ultra here in this use case. It seems like a dual GPU setup on a 7,1 could be better than the M2 Ultra, depending on the codecs you edit I'd imagine?