Chart: 4 M1 SoCs & discussion

[Henry Richardson]

There are now 4 variations of the M1 SoC. Awhile back I made a chart to help me keep track of them while I was trying to decide which Mac to order. I have added the M1 Ultra now. Note, this shows all the various options with each one, but for a particular computer all those options may not exist. I made the chart 2 ways. This first one is easier to read embedded here, but the second one is easier to read if you click on Original to see it larger.





The single core performance is the same for all 4 of these SoCs, but the multi-core performance varies depending on how many cores (and if the software really takes advantage of them). Note that the memory speed though is different so in real life even the single core performance can vary since some programs are making lots more memory accesses than others.

You will note that while the M1 and M1 Pro (one variation) both have 8 CPU cores, the M1 Pro has 2 more performance cores and 2 less efficiency cores.

Awhile back I recall seeing a video or reading somewhere that the M1 computers (M1 Mac Mini, M1 MacBook Air, iMac) have fast SSDs. Faster than the Intel Macs and faster than most Windows PCs. I also saw something about how the M1 Pro, M1 Max, and M1 Ultra computers have even faster SSDs than the M1 computers. Of course, it is nice to have very fast SSDs, but it is particularly important since there is often swapping going on. A very fast SSD can make that swapping much less noticeable. I have also read that the compression and decompression for data in memory is done in hardware and very fast so that also helps a lot. These are 2 of the reasons why one can often get by pretty well with less memory than with the earlier Intel Macs. I am sure there are other reasons as well.

Feel free to correct anything I got wrong and also to add your knowledge to this thread. I am interested in this and find the architecture to be elegant. And that is on top of the change to a RISC CPU.

I suppose later there will be a series of M2 SoCs.

--
Henry Richardson
http://www.bakubo.com

 

[JT26]

What would you like to discuss Henry?

 

[Tom_N]

macOS will do swapping not just to the SSD, but to compressed memory. This started before the switch to Apple Silicon. The hypothesis is that modern CPUs are so fast that the overhead of compression and decompression is preferable to the delays of waiting to shuffle data to and from relatively slow SSDs and HDDs.

The M1-family CPUs might be able to do compression and decompression faster than the Intel CPUs that they replaced.

 

[Ergo607]

I was not aware that RAM was faster along the M1 family. Excellent stuff to put a chart this way. I have bookmarked your thread...

--
Cheers,
Marc

 

[CAcreeks]

There are now 4 variations of the M1 SoC. Awhile back I made a chart to help me keep track of them while I was trying to decide which Mac to order...

What strikes me about this chart is that the only CPU to be possibly faster running DxO Deep PRIME, which uses the neural engine, is the M1 Ultra.

Ever since PhotoLab 5, I've been shocked at the speed of it on my low-end M1 Air.

 

[Tom_N]

There are now 4 variations of the M1 SoC. Awhile back I made a chart to help me keep track of them while I was trying to decide which Mac to order...

What strikes me about this chart is that the only CPU to be possibly faster running DxO Deep PRIME, which uses the neural engine, is the M1 Ultra.

Ever since PhotoLab 5, I've been shocked at the speed of it on my low-end M1 Air.

I believe that in other threads about PL5 and the Neural Engine, people made a guess that DXO might run as well using GPU acceleration on a 16-GPU-core M1 Pro, as it does using Neural Engine acceleration on a M1.

So we might expect relative DXO PL5 DeepPrime speeds of:

• M1 – 1x speed (16-core Neural Engine) or 0.5x speed (8-core GPU)
• M1 Pro – 1x speed (16-core Neural Engine) or 1x speed (16-core GPU)
• M1 Max – 1x speed (16-core Neural Engine) or 2x speed (32-core GPU)
• M1 Ultra – 2x speed (32-core Neural Engine) or 4x speed (64-core GPU)

Caveat: very rough back of the envelope guesses – not confirmed by testing on hardware.

 

[CAcreeks]

I believe that in other threads about PL5 and the Neural Engine, people made a guess that DXO might run as well using GPU acceleration on a 16-GPU-core M1 Pro, as it does using Neural Engine acceleration on a M1. So we might expect relative DXO PL5 DeepPrime speeds of:

• M1 – 1x speed (16-core Neural Engine) or 0.5x speed (8-core GPU)
• M1 Pro – 1x speed (16-core Neural Engine) or 1x speed (16-core GPU)
• M1 Max – 1x speed (16-core Neural Engine) or 2x speed (32-core GPU)
• M1 Ultra – 2x speed (32-core Neural Engine) or 4x speed (64-core GPU)

Caveat: very rough back of the envelope guesses – not confirmed by testing on hardware.

Is it possible to set the processor in DxO PhotoLab? Topaz can do it, but I have not seen any UI for changing the processor in PL5.

In Geekbench 5 measurements, M1 Ultra is almost as fast as AMD Threadripper 3990X. Wow, same price, but Apple delivers the entire rest-of-system for free.

https://www.tomshardware.com/news/apple-m1-ultra-nearly-matches-amd-threadripper-3990x

In the cpubenchmark.net table, only 3995WX and three EPYC chips are faster than 3990X.

 

[Tom_N]

Is it possible to set the processor in DxO PhotoLab? Topaz can do it, but I have not seen any UI for changing the processor in PL5.

I believe that it is, according to the threads in this forum that report that Neural Engine acceleration is faster on the M1. I don't know if PL5 displays the choice if you are running it on an Intel-based Mac. Perhaps someone who has a M1 and PL5 could post a screen shot?

 

[CAcreeks]

Is it possible to set the processor in DxO PhotoLab? Topaz can do it, but I have not seen any UI for changing the processor in PL5.

I believe that it is, according to the threads in this forum that report that Neural Engine acceleration is faster on the M1. I don't know if PL5 displays the choice if you are running it on an Intel-based Mac. Perhaps someone who has a M1 and PL5 could post a screen shot?

I don't remember that discussion in this forum, but you are correct, sir!

 

[QuietOC]

The single core performance is the same for all 4 of these SoCs...

That is rather disappointing. I was expecting the heatsink in the Studio to allow better single thread performance. Intel is looking better again.

 

[Henry Richardson]

Is it possible to set the processor in DxO PhotoLab? Topaz can do it, but I have not seen any UI for changing the processor in PL5.

I believe that it is, according to the threads in this forum that report that Neural Engine acceleration is faster on the M1. I don't know if PL5 displays the choice if you are running it on an Intel-based Mac. Perhaps someone who has a M1 and PL5 could post a screen shot?

I don't remember that discussion in this forum, but you are correct, sir!

If you get a chance it would be cool if you would time it using the 3 choices and let us know the results. Thanks.

--
Henry Richardson

http://www.bakubo.com

 

[Tom_N]

In Geekbench 5 measurements, M1 Ultra is almost as fast as AMD Threadripper 3990X. Wow, same price, but Apple delivers the entire rest-of-system for free.

https://www.tomshardware.com/news/apple-m1-ultra-nearly-matches-amd-threadripper-3990x

The M1 Ultra was 47.8% faster than the AMD Ryzen Threadripper 3990X on the single-threaded benchmark. The M1s in the Mac Minis and MacBook Airs can't keep up with the M1 Ultra or Threadripper for multi-threaded work – but on single-threaded work, they beat the Threadripper by similar margins to the M1 Ultra.

The AMD Ryzen Threadripper 3990X seems to be designed for data center use, where single-core performance is not as important as the ability to spin up many VMs – each running a customer-supplied container that might be more network/database-bound than CPU-bound.

Given the reported power consumption difference, it's fortunate for AMD that the M1 Ultra doesn't run the Intel instruction set, and that Apple is using it in desktop PCs, as opposed to selling it to other systems builders and cloud computing companies who already have a foothold in the data farm hardware market.

 

[CAcreeks]

If you get a chance it would be cool if you would time it using the 3 choices and let us know the results. Thanks.

Looks like CPU only was better than my 7-core (IIRC) GPU. Probably the order mattered, as some PL5 pages were probably already in memory due to MacOS scheduler, which explains why last test was fastest. This GFX 100S image is from a DPreview gallery. Order of tests:

• 36.23 seconds - Apple Neural Engine
• 108.41 - GPU Apple M1
• 107.21 - CPU only
• 31.32 - Auto selection

downsampled because who needs 100Mp?

 

[Tom_N]

The single core performance is the same for all 4 of these SoCs...

That is rather disappointing. I was expecting the heatsink in the Studio to allow better single thread performance. Intel is looking better again.

If Anandtech's article is correct, the CPU core designs are basically the same for all of the M1-family processors. As you move up the line, you get more cores and a higher ratio of performance cores to efficiency cores.

The Core i9-12900K beat the M1 Ultra by 11.4% in single-threaded benchmarks. It lost to it in the multi-threaded ones, where the M1 Ultra was 39.8% faster.

Keep in mind that the Core i9-12900K is a desktop CPU with a TDP of 125 to 241 watts . Yet on single-threaded benchmarks, it is only 17% faster than the M1 in an M1 MacBook Air, a computer that runs for hours off battery power, with no fans.

The list price on the Core i9-12900K is $589. For $699, Apple will sell you a M1 Mac Mini with an entire computer (8 GB RAM, 256 GB SSD, assorted high-speed ports, case, power supply, and operating system) attached. Just bring your own keyboard/mouse/monitor.

What was that you were saying about Intel looking better again?

 

[Henry Richardson]

If you get a chance it would be cool if you would time it using the 3 choices and let us know the results. Thanks.

Looks like CPU only was better than my 7-core (IIRC) GPU. Probably the order mattered, as some PL5 pages were probably already in memory due to MacOS scheduler, which explains why last test was fastest. This GFX 100S image is from a DPreview gallery. Order of tests:

• 36.23 seconds - Apple Neural Engine
• 108.41 - GPU Apple M1
• 107.21 - CPU only
• 31.32 - Auto selection

Thank you for checking. Was this on an M1 MacBook Air 8gb/256gb 8/7? And run on the full size GFX 100S photo?

--
Henry Richardson
http://www.bakubo.com

 

[CAcreeks]

Looks like CPU only was better than my 7-core (IIRC) GPU. Probably the order mattered, as some PL5 pages were probably already in memory due to MacOS scheduler, which explains why last test was fastest. This GFX 100S image is from a DPreview gallery. Order of tests:

• 36.23 seconds - Apple Neural Engine
• 108.41 - GPU Apple M1
• 107.21 - CPU only
• 31.32 - Auto selection

Thank you for checking. Was this on an M1 MacBook Air 8gb/256gb 8/7? And run on the full size GFX 100S photo?

Yes to both questions.

Each time I exited PhotoLab, as it said I should, but I did not reboot. It takes longer to downsample the image, as I did for the post, but not the timings.

IMO PhotoLab isn't particularly good at downsampling. The best algorithm is Bicubic and there is no sharpening. Too bad Irfanview (Lanczos) doesn't run on MacOS. [Added: XnView has Lanczos option.]

I was planning to get a Macbook 16, but I can't complain about 36 seconds for 100Mp! Maybe I'll wait for a newer model with HDMI 2.1.

 

[Henry Richardson]

This is an informative new video.

Mac Studio: 13 Things Apple DIDN'T Tell You at the Event

 

[silentstorm]

The single core performance is the same for all 4 of these SoCs...

That is rather disappointing. I was expecting the heatsink in the Studio to allow better single thread performance. Intel is looking better again.

If Anandtech's article is correct, the CPU core designs are basically the same for all of the M1-family processors. As you move up the line, you get more cores and a higher ratio of performance cores to efficiency cores.

The Core i9-12900K beat the M1 Ultra by 11.4% in single-threaded benchmarks. It lost to it in the multi-threaded ones, where the M1 Ultra was 39.8% faster.

Keep in mind that the Core i9-12900K is a desktop CPU with a TDP of 125 to 241 watts . Yet on single-threaded benchmarks, it is only 17% faster than the M1 in an M1 MacBook Air, a computer that runs for hours off battery power, with no fans.

The list price on the Core i9-12900K is $589. For $699, Apple will sell you a M1 Mac Mini with an entire computer (8 GB RAM, 256 GB SSD, assorted high-speed ports, case, power supply, and operating system) attached. Just bring your own keyboard/mouse/monitor.

What was that you were saying about Intel looking better again?

I think the ultra has a teeny bit better single core performance, from 1753 to 1793. LOL!

Considering the Ultra has 20 cores and the i9-12900K has 16 cores, not surprising that the Ultra wins multi cores score. And it also looks like the geekbench test is not optimised for more threads since Ultra runs 20 and the i9 has 24, also assuming the Ultra runs same as i9 at 3.2GHz base.

Talking about power, the Ultra is not far off actually. The Max eats 60W, and putting 2 together eat 120W at base. And the Max when at boost mode eats about 95W, so that makes it around 190W. And people who want crunching power are not going to care for 250W draw for a cpu. But I have to agree the power efficiency bit, many hours on battery and no fans is a HUGE PLUS! I only wish apple makes the Ultra runs a little faster at 3.5GHz, that should keep many people happy for a very long time.

The price, like you said, is really hard to beat! For $700 you get the whole good performance computer less the KB, mouse and monitor. For the same money, one can only go for i3/Ryzen 3 series system. The only thing Intel/AMD look good is when you need expansion flexibility. There are people who use the i3 to make their own NAS or server for home use, so the flexibility to put a RAID card or extra optical fiber networking card is a must have. Using dongles and protocol converters on the Mini or Studio can be messy and also compatibility problems.

Different strokes for different folks. The Mini and Studio can do so many things really well. :-D

 

[silentstorm]

Awhile back I recall seeing a video or reading somewhere that the M1 computers (M1 Mac Mini, M1 MacBook Air, iMac) have fast SSDs. Faster than the Intel Macs and faster than most Windows PCs. I also saw something about how the M1 Pro, M1 Max, and M1 Ultra computers have even faster SSDs than the M1 computers. Of course, it is nice to have very fast SSDs, but it is particularly important since there is often swapping going on. A very fast SSD can make that swapping much less noticeable. I have also read that the compression and decompression for data in memory is done in hardware and very fast so that also helps a lot. These are 2 of the reasons why one can often get by pretty well with less memory than with the earlier Intel Macs. I am sure there are other reasons as well.

Feel free to correct anything I got wrong and also to add your knowledge to this thread. I am interested in this and find the architecture to be elegant. And that is on top of the change to a RISC CPU.

I suppose later there will be a series of M2 SoCs.

The SSD speed is the same as those PCIe 4 SSD, not any faster in comparison, but is plenty PLENTY fast.

The ram embedded is DDR5 ram running at 5000 speed (from reviews), not the current DDR5 ram sticks on the shelves running at 4800. Besides faster speed, the embedded ram has lower latency as well.

Another thing to note is that the M1 Pro/Max/Ultra has many codecs built-in, not just decoder. For video use, those are very useful and efficient. Neural engines boost for those AI software. This SoC is really a game changer. Can't wait for all apps to be ported to native M1 architecture.

I believe if you have enough ram, the swapping should be min. So getting a 32GB M1 Pro is the min one should buy IMO. Hopefully the new Air comes with basic M1 with 32GB option, that would be enough for many people's use. :-D

 

[joger]

The single core performance is the same for all 4 of these SoCs...

That is rather disappointing. I was expecting the heatsink in the Studio to allow better single thread performance. Intel is looking better again.

If Anandtech's article is correct, the CPU core designs are basically the same for all of the M1-family processors. As you move up the line, you get more cores and a higher ratio of performance cores to efficiency cores.

The Core i9-12900K beat the M1 Ultra by 11.4% in single-threaded benchmarks. It lost to it in the multi-threaded ones, where the M1 Ultra was 39.8% faster.

Keep in mind that the Core i9-12900K is a desktop CPU with a TDP of 125 to 241 watts . Yet on single-threaded benchmarks, it is only 17% faster than the M1 in an M1 MacBook Air, a computer that runs for hours off battery power, with no fans.

The list price on the Core i9-12900K is $589. For $699, Apple will sell you a M1 Mac Mini with an entire computer (8 GB RAM, 256 GB SSD, assorted high-speed ports, case, power supply, and operating system) attached. Just bring your own keyboard/mouse/monitor.

What was that you were saying about Intel looking better again?

could have said it better - perfectly done

I think the ultra has a teeny bit better single core performance, from 1753 to 1793. LOL!

What do teens have to do with benchmarks?

Considering the Ultra has 20 cores and the i9-12900K has 16 cores, not surprising that the Ultra wins multi cores score. And it also looks like the geekbench test is not optimised for more threads since Ultra runs 20 and the i9 has 24, also assuming the Ultra runs same as i9 at 3.2GHz base.

synthetic benchmarks are best case a starting point for further real world examinations.

The most attractive part of the Apple SOCs is the low power consumption for CPU and GPU. As a user I care about the electric bill for my tools and the heat that needs to be pulled away by the AC in my office.

In case in real world I'll have to wait 2 seconds more on a 30 minutes tasks while saving $ 1 on the electrical bill I'd always vote for the lower noise and lower cost option. In fact the M1 studio might be a few seconds faster in many real world tasks while saving power.

Nothing to complain about.

Talking about power, the Ultra is not far off actually. The Max eats 60W, and putting 2 together eat 120W at base. And the Max when at boost mode eats about 95W, so that makes it around 190W. And people who want crunching power are not going to care for 250W draw for a cpu. But I have to agree the power efficiency bit, many hours on battery and no fans is a HUGE PLUS! I only wish apple makes the Ultra runs a little faster at 3.5GHz, that should keep many people happy for a very long time.

People are happy with the M1 baseline and I am very happy with the M1 MAX and I am sure people will be utterly happy with their M1 Ultra for years to come. Optimized software on MACs just shines - FCP is IMHO unbeatable in perceived performance.

the Mac Studio daws maximum 360 W - total system load. That a bit more than those new Intel and AMD chips alone without any GPU

The price, like you said, is really hard to beat! For $700 you get the whole good performance computer less the KB, mouse and monitor. For the same money, one can only go for i3/Ryzen 3 series system. The only thing Intel/AMD look good is when you need expansion flexibility.

What do you want to expand?

Buy the machine as you want it to be and add storage with a NAS. All else is fine as it is and replace your machine in 4..5 years with the next version - you'll be able to sell it for 30..40 % of your purchasing price and when you deduct it to the tax office you'll get in germany 45 % back from the community => 10..20 % real cash out and you'll always have the best system available.

There are people who use the i3 to make their own NAS or server for home use, so the flexibility to put a RAID card or extra optical fiber networking card is a must have. Using dongles and protocol converters on the Mini or Studio can be messy and also compatibility problems.

I am using two NAS - DS1817+ with 8x 10 TB and DS1821+ with 8x 18 TB - both in raid 6 with the old one acting as the backup for the rater new one - both with 10 GbE - hardly any complaint about sped and flexibility - both running cool at a few Watt only - and not in m office - which is almost more important than the speed - i hate fan noise - even the faintest one.

Different strokes for different folks. The Mini and Studio can do so many things really well. :-D

I see nothing these MACs can't do outside of gaming - which is not my goal anyway - I drive real sports and V8 cars

Joking aside - the new studio looks amazing and I am tempted to get one just for fun - it's a bargain in investment compare to it's computational power.