Mac Studio M1 Ultra 48 core vs 12900/3090

[darthaddie]

Hi. Just got my Mac Studio M1 Ultra from the Apple store. comparing it to my 12900/309o system for Photo and video editing as I write.

PC Specs:

Intel 12900k overclocked to 5 Ghz all core (3.9Ghx E cores) - water-cooled
DDR5 64GB 5200 Mhz RAM
Asus RTX 3090 water-cooled
MSI Z690 ACE motherboard
8 TB NVME RAID

Mac Specs:
M1 Ultra
20 Core CPU
48 Core GPU
64GB RAM

Both are connected to EIZO 4k displays separately.

I just ran a Lightroom classic export of 1000 images from my recent wedding shoot. Here are the results. I am running DaVinci Studio benchmarks on my recent wedding edit. Its a very heavy edit with Temporal sharpening, color grade and halation effects. So the results will be a good comparison.

Lightroom export 1000 RAW images from my SL2/SL2S camera (Approx 50% mix), Standard Sharpening, Rename files, 90% JPEG. No Metadata changes

Time taken by PC - 20:39 minutes
Mac Studio: 16:33 minutes

Power consumed by PC - 250-380W
Mac Studio: 60W (lol)

Other things to note:
Lightroom on PC has an annoying flicker while exporting. This is with all PC's

I will post back with Davinci results. I am really interested in checking the GPU bound effects vs the 3090.

 

[Lone Deranger]

That's a great performance for the Ultra. Thank you for posting this comparison.

Would you be able/willing to run some Blender benchmark tests?

Demo Files — blender.org

Home of the Blender project - Free and Open 3D Creation Software

www.blender.org

The 3090 still seems well ahead in terms of GPU based rendering (Eevee and Cycles) compared to the Ultra.

 

[Jorbanead]

That's a great performance for the Ultra. Thank you for posting this comparison.

Would you be able/willing to run some Blender benchmark tests?

Demo Files — blender.org

Home of the Blender project - Free and Open 3D Creation Software

www.blender.org

The 3090 still seems well ahead in terms of GPU based rendering (Eevee and Cycles) compared to the Ultra.

Blender is not at all optimized for apple silicon, so I don’t think it’s really a fair comparison if that’s what you’re going for. It "runs" on Apple Silicon, but it's still using Optix rendering engine, not metal. So currently it's going Optix to open CL to Metal. Apple is working on an Optix to Metal converter, and it will be built into the OS. Once that happens, it'll be a lot faster.

 

[dmr727]

Blender is not at all optimized for apple silicon, so I don’t think it’s really a fair comparison if that’s what you’re going for.

I'd argue it's fair for someone that uses their machine for Blender work. But to your point, lots of people are using benchmarks from tools they don't actually use just to justify an argument (or troll) they're trying to make. That Blender benchmark is useless for anyone that's not actually using Blender, such as the OP using Lightroom.

 

[Matt2012]

Can you please do a test in Davinci with some temporal noise reduction and see how it compares to the 3090?
I was dissapointed with the top spec M1 max laptop in that test so hoping Studio performs faster.

 

[Pressure]

I'd argue it's fair for someone that uses their machine for Blender work. But to your point, lots of people are using benchmarks from tools they don't actually use just to justify an argument (or troll) they're trying to make. That Blender benchmark is useless for anyone that's not actually using Blender, such as the OP using Lightroom.

Apple first joined the Blender Development Fund 6 months ago. I wouldn't expect miracles in that timeframe.

Redshift on the other hand seems to be quite solid.

 

[Jorbanead]

I'd argue it's fair for someone that uses their machine for Blender work. But to your point, lots of people are using benchmarks from tools they don't actually use just to justify an argument (or troll) they're trying to make. That Blender benchmark is useless for anyone that's not actually using Blender, such as the OP using Lightroom.

Agreed 100%. I just worry that most people are doing these blender benchmarks simply just to see how the GPU’s compare, and when they see that blender isn’t that great, they immediately think the GPU on the ultra is terrible. I’ve already seen it blasted all over FB groups.

I agree Apple may have been a been to arrogant in their GPU claims, but I also want to play fair too.

 

[darthaddie]

UPDATE: I am seeing Davinci suffer badly when you apply Temporal noise reduction.

Without any NR the 4K60 10bit 4:2:0 footage exports at 170 fps which is crazy as the 3090 maxes out at 90-100 fps.

With the NR applied however the export times are about 35-40 fps with the Ultra and 70 fps on the 3090.

I am thinking the CPU is a lot faster along with the Media engines but the GPU is lagging. Timeline performance with the NR applied also slows down the performance noticeably. A 1/2 resolution fixes that.

UPDATE 2: I monitored the GPU usage under each export on the Ultra and only 30-40% is being used. CPU use is around 50%. Can't check the media engines though. I am pretty sure that an update is required.

UPDATE 3: The only slow down is the Temporal Noise reduction. LUTs and Color grade have little effect. There are some Davinci transitions which work worse on the M1 Ultra compared to a 3090.

For those wondering, 4K 60fps 10bit 4:2:2 with color grade (Color Transform and Kodak Film LUT applied) from the Sony FX3 or 4k60fps 10bit 4:2:0 from the Panasonic S5 render at 100% speed gain when compared to a 3090. the 3090 gets 70fps and the M1 Ultra over 150fps.

Update 4: During a really heavy 4k export with color grading, noise reduction applied the 3090 PC consumes 5750w - 650w. The M1 Ultra is at 105W-125W. Its almost laughable. Timeline scrubbing is out of this world good. both with H265 and 264. I tried 4k from the Panasonic, R5 and the Sony FX3. Either in 10 bit 4:2:2 or 10 bit 4:2:0. all at 60fps.

The 8K Canon R5 RAW Lite footage plays back just fine. the 3090 struggles a bit. Not a very big difference but still noticeable.

Another important thing and this probably needs a Mac OS / Apps update. Mac OS randomly becomes sluggish when using any of the NLE's. Same with Premiere Pro and Davinci. Giving it a couple of secs, revives the performance.

Update 5: Blender is a little bit slower compares to a 3090. But if the 64 core GPU scales well it will almost match the performance. Viewports are really good though. Not 3090 good, but real close.

I see terrible comments from people on Reddit, where I posted this first (mostly PC enthusiasts) that I was just trying to shill the Mac. lol. I am not. I have both systems and each one servers its purpose. The PC is now used mostly for gaming as I love working on the Mac with my edits.

To summarize:

• Davinci performs 50-100% faster depending on what you're doing in most tasks. and sometimes 50% slower (Noise reduction). Face tracking and some other effects don't seem to slow down the timeline.
• Lightroom is wicked fast with editing and exports. Beats the 12900k hands down
• FCPX timeline performance is super awesome. FCP exports to H.265 (the other thing I tested are stupid slow). I don't know why. a 20 Sec 4k clip export to H.265 took 4 mins. Its almost mind boggling. But I don't use FCP X much.
• Not disappointed at all. :-D

 

[PianoPro]

Appreciate getting your real world performance. But do you have any idea what is going on where it appears the Ultra GPU seems to be constrained to much lower power levels and performance on the graphics benchmarks in this video?

 

[darthaddie]

Appreciate getting your real world performance. But do you have any idea what is going on where it appears the Ultra GPU seems to be constrained to much lower power levels and performance on the graphics benchmarks in this video?

Whoa! This probably and answers my question of why the entire system would randomly slow down for no reason. Like on 20GB RAM used, CPU at 40%, GPU at 20% and the Mac feels sticky.

Could Apple be holding back the performance because of the power limitations? Or do they need to release an update to fix this like the MacBook Pro earlier.

Or are they not wanting to fix this to force upgrade to the next Mac OS.

Oh!!!! I don't want the Mac to be held back by this issue. Its got lots off thermal headroom....

 

[ctjack]

Its got lots off thermal headroom....

Actually it doesn't. M1 sitting fanless in Airs were fantastic - mine sits at idle browsing at 27C which is ridiculous. But the problem predicted and that eventually revealed - you can increase the power of the M1 as in MBP 13/Mini, M1 Pro/Max but it will run much hotter than it will give gains. In other words, increasing the performance by 2x gets you 3-4x hotter CPU - yeah, no magic here as the people were referring exactly to this because well physics is physics.

 

[darthaddie]

Actually it doesn't. M1 sitting fanless in Airs were fantastic - mine sits at idle browsing at 27C which is ridiculous. But the problem predicted and that eventually revealed - you can increase the power of the M1 as in MBP 13/Mini, M1 Pro/Max but it will run much hotter than it will give gains. In other words, increasing the performance by 2x gets you 3-4x hotter CPU - yeah, no magic here as the people were referring exactly to this because well physics is physics.

That is a good point. But then where’s the 2x or 1.5x performance compared to a M1 Max. It’s almost the same. The Mac studio ultra in that case doesn’t justify the cost or the claims that Apple did.

 

[mi7chy]

Update 5: Blender is a little bit slower compares to a 3090.

I see terrible comments from people on Reddit, where I posted this first that I was just trying to shill the Mac.

How is 4x to almost 5x difference just a little bit?

 

[PianoPro]

Actually it doesn't. M1 sitting fanless in Airs were fantastic - mine sits at idle browsing at 27C which is ridiculous. But the problem predicted and that eventually revealed - you can increase the power of the M1 as in MBP 13/Mini, M1 Pro/Max but it will run much hotter than it will give gains. In other words, increasing the performance by 2x gets you 3-4x hotter CPU - yeah, no magic here as the people were referring exactly to this because well physics is physics.

No. The fans don't even ramp up above idle and the temperature is very low. There's a ton of headroom it just isn't getting used.

 

[ctjack]

No. The fans don't even ramp up above idle and the temperature is very low. There's a ton of headroom it just isn't getting used.

Hopefully so. Otherwise could be it is low and cold because the CPU is being held back.

 

[darthaddie]

How is 4x to almost 5x difference just a little bit?

I don’t use blender. But I ran the Party tug scene with Metal CPU/GPU turned on with EAVEE (don’t know if I got that right).

M1 ultra rendered in 4 seconds and the 3090 in 2.5 secs. I guess that was not 4x slower?

As I don’t use it, I cannot confirm what this means in real world use.

 

[iPadified]

UPDATE: I am seeing Davinci suffer badly when you apply Temporal noise reduction.

Without any NR the 4K60 10bit 4:2:0 footage exports at 170 fps which is crazy as the 3090 maxes out at 90-100 fps.

With the NR applied however the export times are about 35-40 fps with the Ultra and 70 fps on the 3090.

I am thinking the CPU is a lot faster along with the Media engines but the GPU is lagging. Timeline performance with the NR applied also slows down the performance noticeably. A 1/2 resolution fixes that.

UPDATE 2: I monitored the GPU usage under each export on the Ultra and only 30-40% is being used. CPU use is around 50%. Can't check the media engines though. I am pretty sure that an update is required.

UPDATE 3: The only slow down is the Temporal Noise reduction. LUTs and Color grade have little effect. There are some Davinci transitions which work worse on the M1 Ultra compared to a 3090.

For those wondering, 4K 60fps 10bit 4:2:2 with color grade (Color Transform and Kodak Film LUT applied) from the Sony FX3 or 4k60fps 10bit 4:2:0 from the Panasonic S5 render at 100% speed gain when compared to a 3090. the 3090 gets 70fps and the M1 Ultra over 150fps.

Update 4: During a really heavy 4k export with color grading, noise reduction applied the 3090 PC consumes 5750w - 650w. The M1 Ultra is at 105W-125W. Its almost laughable. Timeline scrubbing is out of this world good. both with H265 and 264. I tried 4k from the Panasonic, R5 and the Sony FX3. Either in 10 bit 4:2:2 or 10 bit 4:2:0. all at 60fps.

The 8K Canon R5 RAW Lite footage plays back just fine. the 3090 struggles a bit. Not a very big difference but still noticeable.

Another important thing and this probably needs a Mac OS / Apps update. Mac OS randomly becomes sluggish when using any of the NLE's. Same with Premiere Pro and Davinci. Giving it a couple of secs, revives the performance.

Update 5: Blender is a little bit slower compares to a 3090. But if the 64 core GPU scales well it will almost match the performance. Viewports are really good though. Not 3090 good, but real close.

I see terrible comments from people on Reddit, where I posted this first (mostly PC enthusiasts) that I was just trying to shill the Mac. lol. I am not. I have both systems and each one servers its purpose. The PC is now used mostly for gaming as I love working on the Mac with my edits.

To summarize:

• Davinci performs 50-100% faster depending on what you're doing in most tasks. and sometimes 50% slower (Noise reduction). Face tracking and some other effects don't seem to slow down the timeline.
• Lightroom is wicked fast with editing and exports. Beats the 12900k hands down
• FCPX timeline performance is super awesome. FCP exports to H.265 (the other thing I tested are stupid slow). I don't know why. a 20 Sec 4k clip export to H.265 took 4 mins. Its almost mind boggling. But I don't use FCP X much.
• Not disappointed at all. :-D

This is how benchmarking should be done! The glittery YouTubers productions are not believable because of an hidden agenda promoting either PC or Mac.

 

[-narcan-]

I don’t use blender. But I ran the Party tug scene with Metal CPU/GPU turned on with EAVEE (don’t know if I got that right).

M1 ultra rendered in 4 seconds and the 3090 in 2.5 secs. I guess that was not 4x slower?

As I don’t use it, I cannot confirm what this means in real world use.

Could you try one of the Cycles demo files (with Blender 3.1 GPU Metal rendering enabled) ? Cycles is the ray tracing renderer.

 

[ctjack]

This is how benchmarking should be done! The glittery YouTubers productions are not believable because of an hidden agenda promoting either PC or Mac.

At this moment all of this back and forth videos don't do any better. Apple with ARM is becoming the next Tesla. Some hate tesla and some love it. Well we all know the pros and cons it really well: low range when compared to usual cars, but much higher torque because of the physics of electric motors(can pull out your truck), then fast acceleration because of this torque, save on brake pads because of recuperative braking, nice and smooth steering wheel adjustment, best in the class infotainment( can play games), best autopilot/lane tracking software, smoothest steering wheel adjustments, lower quality interior.

The same applies to Apple. It is incredibly fast in some things and incredibly bad at others(gaming or some unoptimized software or even optimized(claims to be so) and less performance).

 

[iPadified]

Could you try one of the Cycles demo files (with Blender 3.1 GPU Metal rendering enabled) ? Cycles is the ray tracing renderer.

Here is a lengthy thread about the subject

3D Rendering on Apple Silicon, CPU&GPU

This is a thread for discussing everything about 3D rendering on Apple Silicon, M1 and beyond. Whether you're using a dedicated CPU renderer like Keyshot, a plugin like Redshift, Cycles, Arnold or Octane for C4D, Maya or Houdini, or even if you're rendering inside ZBrush with its native...

forums.macrumors.com

 

[mi7chy]

I don’t use blender. But I ran the Party tug scene with Metal CPU/GPU turned on with EAVEE (don’t know if I got that right).

M1 ultra rendered in 4 seconds and the 3090 in 2.5 secs. I guess that was not 4x slower?

As I don’t use it, I cannot confirm what this means in real world use.

EEVEE isn't the primary method of Blender 3D rendering and only uses legacy OpenGL which doesn't maximize the capability of GPU compared to Cycles X.

 

[iPadified]

EEVEE isn't the primary method of Blender 3D rendering and only uses legacy OpenGL which doesn't maximize the capability of GPU compared to Cycles X.

Hardly old tech. It was introduced in 2.8. It is fast irrespective of hardware but the output is not as good as cycles.

 

[mi7chy]

Fortunately, there are more competent and unbiased comparisons out there. Currently watching this one.

 

[hugodrax]

Wow thats crazy how many watts the PC burns.

 

[l0stl0rd]

EEVEE isn't the primary method of Blender 3D rendering and only uses legacy OpenGL which doesn't maximize the capability of GPU compared to Cycles X.

Yes that is why they started rewriting Eevee it will be Metal in the case of Apple (hopefully sometime this year).

EEVEE Next

Do I get this statement right that at the moment EEVEE Next (in master) does not render anything at all, because nothing has been implemented yet? No critique, just curious if that’s what you’re saying.

devtalk.blender.org