M3 Ultra Benchmarking

[v3rlon]

I have my M3 Ultra with 60 core GPU. I set up with "restore from Time Machine backup" that dates back multiple computers into the Intel MacBook Pro era. Yeah, but a cold install of everything is a pain. It might just be time though.
Anyway. I fired up Geekbench, and here is what I have. Note that this is not some hyper optimized setup. This is me running the computer on startup like a normal day in the life, but it is a fresh boot.
Single Core 3257
Multicore 27551
OpenCL. 129134
Metal. 231223

Still, the multicore and compute scores look a little low to me. This has me thinking a 'format the system drive and start over clean,' is about due. Does anyone else have benchmark results?

 

[Basic75]

the multicore and compute scores look a little low to me

Why, are they lower than other M3 Ultra machines?

 

[Basic75]

Geekbench Search - Geekbench

 

[v3rlon]

Why, are they lower than other M3 Ultra machines?

That is my big question. When I posted it, Geekbench didn’t have any other M3ultra listings.

 

[M4pro]

Geekbench results vary - a lot -from run to run.

It’s fair to question how much GB scores actually relate to a user’s own system.

Since the company has never explained the discrepancy in scores, they difference could just as likely relate to the amount of network traffic at any moment on a users internet connection.

 

[v3rlon]

Geekbench results vary - a lot -from run to run.

It’s fair to question how much GB scores actually relate to a user’s own system.

Since the company has never explained the discrepancy in scores, they difference could just as likely relate to the amount of network traffic at any moment on a users internet connection.

I did use more than one run.
I am pretty sure the Geekbench score is a measure of how fast the system can run Geekbench (as opposed to whatever real applications they run). It shouldn't be about network traffic at all.

I would still expect a faster computer (M3 Ultra) to be faster that a slower one (M2 Ultra). While a 50% faster Geekbench score may not reflect a 50% faster video encode score or a 50% higher frame rate in Baldur's Gate 3, It is a starting point, and these things have to be measured somehow. Not everyone cares how fast something renders in Lightwave or uses my exact workflow in Affinity Photo, but how do you judge a computer without the benchmarks and without just buying it and setting it up and testing it?

 

[M4pro]

I did use more than one run….It shouldn't be about network traffic at all…

If you did multiple runs - then GB returned multiple varying results.

Stangely, the developer hasn’t been upfront about the discrepancy in results everyone experiences.

Given that, network traffic is just as likely as any other possible cause.

 

[crazy dave]

I have my M3 Ultra with 60 core GPU. I set up with "restore from Time Machine backup" that dates back multiple computers into the Intel MacBook Pro era. Yeah, but a cold install of everything is a pain. It might just be time though.
Anyway. I fired up Geekbench, and here is what I have. Note that this is not some hyper optimized setup. This is me running the computer on startup like a normal day in the life, but it is a fresh boot.
Single Core 3257
Multicore 27551
OpenCL. 129134
Metal. 231223

Still, the multicore and compute scores look a little low to me. This has me thinking a 'format the system drive and start over clean,' is about due. Does anyone else have benchmark results?

Here are 80-core results:

Apple’s new Mac Studio delivers stellar performance in M3 Ultra form

Apple’s new Mac Studio doesn’t look different on the outside, but there are plenty of performance upgrades under the hood

www.tomshardware.com

I also saw metal scores for the 80-core GPU of 263,000. So without a lot of other bind Ultra variants against to compare, I'd say a priori yours looks fine.

If you did multiple runs - then GB returned multiple varying results.

Stangely, the developer hasn’t been upfront about the discrepancy in results everyone experiences.

Given that, network traffic is just as likely as any other possible cause.

It's not? GB 6 doesn't rely on network traffic at all. Everything is run locally. The free version only works with an internet connection to make sure the result is posted which is how they sell the paid to version to companies and reviewers who don't want their results posted, especially before release. But yeah the test works just fine offline (even the free version, it just won't tell you the result).

From what I can tell, just about every benchmark varies by at least 5% on a single machine and about 10% between machines. GB does tend to be worse than that, but the culprit is likely how short individual subtests are. This means it's heavily reliant on the transient boost clocks achieved by the processor during the test which is very temperamental especially between machines but even within a single machine - especially on x86 processors but even on Apple. For example: I remember for the GB 5 GPU test, the M1 Ultra had pretty bad scaling relative to the Max, but the GB 5 GPU scaling was atrociously bad and Andrei and Ryan at Anandtech basically confirmed that the problem was GB 5's runs were so short that Apple's slow clock ramping on the Ultra in particular meant that the GPU was never hitting its max clocks during the individual tests. GB 6 fixed that particular issue thankfully, but overall for both the GPU and CPU the tests are still very short.

 

[theorist9]

I thought this post was informative. At least for rendering with Redshift and Cinema 4D, a 60-core GPU M3 Ultra is significantly slower than a 4070 Ti Super/14900KF, in spite of the Ultra's faster CPU. I don't know how much of that is the relative optimization of that software for AS vs PC, and how much of it is the GPU's themsevles.

I'm one of those few people who have a 2019 Mac Pro with maxed out MPX modules and just recently a base 2025 Studio M3 Ultra. I've was using the Mac Pro for 3D rendering and whilst I have a PC with a couple of Nvidia cards and it's quicker than both of the Macs, I just prefer using Mac OS.

The fact is those, this studio runs rings around the MacPro for straight performance. It also generates way less heat and is silent, noting my PC can claim. Yeah there are still some speciality audio cards but the majority of audio interfaces are still USB-C, there are only a few Thunderbolt interfaces. PCIe audio are pretty scarce.

Blow are my results. Unless they come out with some M4/M5 Ultra/Extreme chip with PCIe slots I can't see the point of it as much I hate to say it.

2019 MacPro with 2x 6800XT Duo's
Self Build PC with 1914900KF + 4070 Ti Super + A4000
Mac Studio M3 Ultra Base 28cpu/60gpu

Redshift Benchmark:
MacPro: 2:30
Studio: 2:03
PC: 1:37

Cinebench Multicore CPU
MacPro : 1169
Studio: 2666
PC: 2016

Rendering an actual scene in Cinema 4D and Redshift. The scene has a lot of translucent objects, think trees, grass etc so not the easiest thing to render. Output was at 3000px X 3000px.

Mac Studio 19:12 RTX ON
Mac Pro 2019 41:48
Windows 11 + 4070 Ti Super 14:58
Windows 11 + 4070 Ti Super + A4000 10:21

 

[v3rlon]

If you did multiple runs - then GB returned multiple varying results.

Stangely, the developer hasn’t been upfront about the discrepancy in results everyone experiences.

Given that, network traffic is just as likely as any other possible cause.

I DID do multiple runs and they were all in this neighborhood.
It just seemed like listing all of them would be redundant, repetitive, boring, pointless, monotonous, and tedious, but maybe that is just me.
Maybe your MIDI interface is acting up because you don’t have enough SIMM modules in your digital computer.

And the faq on their website does cover a few things that impact scores without giving away the secret sauce. Because all the major players, including Intel, nvidia, AMD, ATI, and others have been caught optimizing for popular Benchmarks, I do understand this

 

[smckenzie]

I thought this post was informative. At least for rendering with Redshift and Cinema 4D, a 60-core GPU M3 Ultra is significantly slower than a 4070 Ti Super/14900KF, in spite of the Ultra's faster CPU. I don't know how much of that is the relative optimization of that software for AS vs PC, and how much of it is the GPU's themsevles.

Thanks.

Here's another set. This is based on an actual scene from Disney's Moana converted for Redshift, its a bit of a monster.

M2 Ultra Mac Studio 128GB ram 512 blocksize: 26:27
M3 Ultra Mac Studio 256GB ram 80c GPU 512 blocksize: 16:35
M3 Ultra Mac Studio 96GB ram 60c GPU 512 blocksize : 20m 5s
2019 MacPro 2x 6800XT Duos: 46:16s

Another scene which a particle simulation, DOF, Subsurface Scattering, Transmission, so not to easy:

M4 Pro 64GB Ram: 11:57
M3 Ultra 60 core GPU 96GB Ram: 11m:28s
2019 MacPro 2x 6800XT Duos: 9m 53s
Win 5950x 128GB w/ 4090: 3m 44s
Win I9 14900KF + 4070 Ti Super: 6m: 27s

 

[theorist9]

Thanks.

Here's another set. This is based on an actual scene from Disney's Moana converted for Redshift, its a bit of a monster.

M2 Ultra Mac Studio 128GB ram 512 blocksize: 26:27
M3 Ultra Mac Studio 256GB ram 80c GPU 512 blocksize: 16:35
M3 Ultra Mac Studio 96GB ram 60c GPU 512 blocksize : 20m 5s
2019 MacPro 2x 6800XT Duos: 46:16s

Another scene which a particle simulation, DOF, Subsurface Scattering, Transmission, so not to easy:

M4 Pro 64GB Ram: 11:57
M3 Ultra 60 core GPU 96GB Ram: 11m:28s
2019 MacPro 2x 6800XT Duos: 9m 53s
Win 5950x 128GB w/ 4090: 3m 44s
Win I9 14900KF + 4070 Ti Super: 6m: 27s

Any idea why the 2019 MP (which has been slower than the M3 Ultra on all your other tests) beat it on the last test?

 

[smckenzie]

Any idea why the 2019 MP (which has been slower than the M3 Ultra on all your other tests) beat it on the last test?

No idea. It the same time on a different Mac Pro so it wasn't a fluke. The 6800XT's struggle on scenes with lots of translucency, caustics and stuff like that because of no RTX. On other simpler scenes its much closer or faster.

 

[M4pro]

But yeah the test works just fine offline (even the free version, it just won't tell you the result)…

From what I can tell, just about every benchmark varies by at least 5% on a single machine and about 10% between machines. GB does tend to be worse than that, but the culprit is likely how short individual subtests are. This means it's heavily reliant on the transient boost clocks achieved by the processor during the test which is very temperamental especially between machines but even within a single machine - especially on x86 processors but even on Apple. For example: I remember for the GB 5 GPU test, the M1 Ultra had pretty bad scaling relative to the Max, but the GB 5 GPU scaling was atrociously bad and Andrei and Ryan at Anandtech basically confirmed that the problem was GB 5's runs were so short that Apple's slow clock ramping on the Ultra in particular meant that the GPU was never hitting its max clocks during the individual tests. GB 6 fixed that particular issue thankfully, but overall for both the GPU and CPU the tests are still very short.

Huh, what? Displaying the results is the whole point.

I haven’t seen other benchmarks be high-variance the way Geekbench is.

Which is an obvious red flag.

 

[crazy dave]

Huh, what? Displaying the results is the whole point.

Yes, not displaying offline results is how they get people to pay for the full version. Displaying the results is not calculating the results. Which part of the score exactly do you think is using a network connection and why?

I haven’t seen other benchmarks be high-variance the way Geekbench is.

Which is an obvious red flag.

Red flag for what? I just explained why the variance is so high. Your insistence that it is network connection doesn't appear to be founded on anything - if it were true it would be easily shown by reviewers running it offline (same program just with license put in and that can be run in either mode, offline or online, which is why GB results often leak by the reviewer manufacturer forgetting to flip the switch) and getting different results never mind everyone on faster vs slower connections getting different results. As an explanation, it just doesn't make sense.

==============

BTW for those who are interested. Geekbench documentation is here: https://www.geekbench.com/doc/geekbench6-benchmark-internals.pdf. The benchmark is not open source I grant you and there a few things I wish they talked about, but there is as an expected "IPC" measurement for the base CPU as well as as other metrics.

 

[crazy dave]

Thanks.

Here's another set. This is based on an actual scene from Disney's Moana converted for Redshift, its a bit of a monster.

M2 Ultra Mac Studio 128GB ram 512 blocksize: 26:27
M3 Ultra Mac Studio 256GB ram 80c GPU 512 blocksize: 16:35
M3 Ultra Mac Studio 96GB ram 60c GPU 512 blocksize : 20m 5s
2019 MacPro 2x 6800XT Duos: 46:16s

Another scene which a particle simulation, DOF, Subsurface Scattering, Transmission, so not to easy:

M4 Pro 64GB Ram: 11:57
M3 Ultra 60 core GPU 96GB Ram: 11m:28s
2019 MacPro 2x 6800XT Duos: 9m 53s
Win 5950x 128GB w/ 4090: 3m 44s
Win I9 14900KF + 4070 Ti Super: 6m: 27s

I'm assuming except for the CB R24 multicore all of these are GPU render times?

I thought this post was informative. At least for rendering with Redshift and Cinema 4D, a 60-core GPU M3 Ultra is significantly slower than a 4070 Ti Super/14900KF, in spite of the Ultra's faster CPU. I don't know how much of that is the relative optimization of that software for AS vs PC, and how much of it is the GPU's themsevles.

Assuming all of this is GPU rendering (except the CB multicore result), then some, but definitely not all, of it is indeed software. As far as I can tell looking at a bunch of GPU rendering benchmarks, the open source Blender has the best GPU renderer for Apple Silicon with Apple having contributed code to it directly. That said, in that best case scenario of Blender 4.3 rendering the 60-core M3 Ultra is still about equivalent to a 4070 Ti and is behind the Super variant (keeping in mind that even these median results may not represent "stock" configurations of said Nvidia GPUs) with the full M3 Ultra just ahead of said Super variant. Given the results from @smckenzie, I doubt the full M3 Ultra would beat his 4070 Ti Super in the Redshift benchmarks (though I'm also assuming that the 4070 Ti Super would simply choke on the Moana benchmark?).

Blender - Open Data

Blender Open Data is a platform to collect, display and query the results of hardware and software performance tests - provided by the public.

opendata.blender.org

 

[smckenzie]

I'm assuming except for the CB R24 multicore all of these are GPU render times?


Assuming all of this is GPU rendering (except the CB multicore result), then some, but definitely not all, of it is indeed software. As far as I can tell looking at a bunch of GPU rendering benchmarks, the open source Blender has the best GPU renderer for Apple Silicon with Apple having contributed code to it directly. That said, in that best case scenario of Blender 4.3 rendering the 60-core M3 Ultra is still about equivalent to a 4070 Ti and is behind the Super variant (keeping in mind that even these median results may not represent "stock" configurations of said Nvidia GPUs) with the full M3 Ultra just ahead of said Super variant. Given the results from @smckenzie, I doubt the full M3 Ultra would beat his 4070 Ti Super in the Redshift benchmarks (though I'm also assuming that the 4070 Ti Super would simply choke on the Moana benchmark?).

Blender - Open Data

Blender Open Data is a platform to collect, display and query the results of hardware and software performance tests - provided by the public.

opendata.blender.org

Yes, all Redshift GPU scores.

I did test 2025 Cinebench MC scores on all 3 machines:

Ultra 28 core: 2666
14900KF: 2011
Xeon 16 core (7,1): 1160 (I think, somewhere around there)

As for the Redshift benchmark:

Ultra 60 core: 2:03
Ultra 80 core: 1:50
Mac Pro 2x 6900 XT Duo: 2:30
4070 Ti Super (ASUS) : 1:37 (I think)

 

[crazy dave]

Yes, all Redshift GPU scores.

I did test 2025 Cinebench MC scores on all 3 machines:

Ultra 28 core: 2666
14900KF: 2011
Xeon 16 core (7,1): 1160 (I think, somewhere around there)

As for the Redshift benchmark:

Ultra 60 core: 2:03
Ultra 80 core: 1:50
Mac Pro 2x 6900 XT Duo: 2:30
4070 Ti Super (ASUS) : 1:37 (I think)

Awesome, thanks! Out of curiosity, what happens if you try to run the Moana benchmark on the 4070 Ti Super? Does it run at all or crash? Or does it just run super slowly? How many GB of VRAM does the Moana render take?

 

[smckenzie]

I had issues running that on the 4070 - just was taking ages doing prep work, something was off with it for sure. I never completed it.