Mac Studio Teardown Reveals Huge M1 Ultra Chip That Dwarfs AMD Ryzen CPU in Size

Tsing

The FPS Review
Staff member
Joined
May 6, 2019
Messages
11,067
Points
83
Apple may have overestimated the performance of its Mac Studio, but if there's one thing that the company didn't exaggerate, it's the size of its M1 Ultra chip, according to a teardown of the system shared by Max Tech's Vadim Yuryev that provides a clear look at how massive the SoC actually looks in person.

Go to post
 
Surprise surprise. I bet peak power consumption is high too.
 
No surprises really. Remember the GPU, system controller, and DRAM are also in that package. The only thing not in there is the power regulation and the SSD. So if you really wanted to compare to a PC you’d need to look at the entire motherboard, not just the CPU.

Apple also disclosed the Ultra’s TDP to be 60W. Not unbelievable really - the M1 Max MBP was measured by Anandtech and pulled peak 40W from the wall during their benchmark runs. I haven’t seen a similar test for the Studio yet.
 
No surprises really. Remember the GPU, system controller, and DRAM are also in that package. The only thing not in there is the power regulation and the SSD. So if you really wanted to compare to a PC you’d need to look at the entire motherboard, not just the CPU.

Apple also disclosed the Ultra’s TDP to be 60W. Not unbelievable really - the M1 Max MBP was measured by Anandtech and pulled peak 40W from the wall during their benchmark runs. I haven’t seen a similar test for the Studio yet.

I am not unimpressed by the studio systems. Honestly if I had the cash to burn I would totally buy one to play with. Hummm... maybe I can convince work to get me one... what do you all think? Many of our software developers use mac book laptops... hummm.
 
I am not unimpressed by the studio systems. Honestly if I had the cash to burn I would totally buy one to play with. Hummm... maybe I can convince work to get me one... what do you all think? Many of our software developers use mac book laptops... hummm.
I've owned several Mac Minis. Those machines are awesome, and I've used them as everything from a HTPC to a mail server to industrial controller. Just the right size where it can pack some power and not be too expensive, and small enough that you can toss it behind the screen or tuck it out of the way. The Studio is a bit more of an investment though, and a bit taller.

I can also say the regular M1 is no slouch. Having ... 6 (+/- ?) of them ... might just be overkill. Unless, of course, you are actually crunching 8K video or training AIs or doing something real. Yeah, I would love to own a studio - but I have no idea what I would do with it. Kinda like a boat - I'd love to own one, but no idea when I'd ever get it out on the water.

I have a M1 13" MBP. I can say hands down it's the best laptop I've ever owned. I wouldn't call it the fastest computer I've ever owned, but it's snappy, gets great battery life, doesn't get hot or loud, feels extremely solid and well built, and runs Office, email, web browsers, and has a native Unix terminal -- all the things I really look for in a laptop

I do have two reservations about it though. I wish I had the 16", coming from a long line of 15" I have a really hard time adjusting to the smaller screen. That, and I have a large library of x86 virtual machines that I use for various things at work that I can't run on it any longer - I have a cheapo $350 laptop just for the purpose of running those tools when I absolutely need them, which is unfortunate, but not a frequent occurrence. I need to get around to trying ARM Windows and see if that software will run there, but haven't fooled with it.
 
I've owned several Mac Minis. Those machines are awesome, and I've used them as everything from a HTPC to a mail server to industrial controller. Just the right size where it can pack some power and not be too expensive, and small enough that you can toss it behind the screen or tuck it out of the way. The Studio is a bit more of an investment though, and a bit taller.

I can also say the regular M1 is no slouch. Having ... 6 (+/- ?) of them ... might just be overkill. Unless, of course, you are actually crunching 8K video or training AIs or doing something real. Yeah, I would love to own a studio - but I have no idea what I would do with it. Kinda like a boat - I'd love to own one, but no idea when I'd ever get it out on the water.

I have a M1 13" MBP. I can say hands down it's the best laptop I've ever owned. I wouldn't call it the fastest computer I've ever owned, but it's snappy, gets great battery life, doesn't get hot or loud, feels extremely solid and well built, and runs Office, email, web browsers, and has a native Unix terminal -- all the things I really look for in a laptop

I do have two reservations about it though. I wish I had the 16", coming from a long line of 15" I have a really hard time adjusting to the smaller screen. That, and I have a large library of x86 virtual machines that I use for various things at work that I can't run on it any longer - I have a cheapo $350 laptop just for the purpose of running those tools when I absolutely need them, which is unfortunate, but not a frequent occurrence. I need to get around to trying ARM Windows and see if that software will run there, but haven't fooled with it.
Why can't you run x86 vm's on the macbook? Is the virtual PC software (that I forget the name of) not capable of emulation on the MBP yet?
 
Why can't you run x86 vm's on the macbook? Is the virtual PC software (that I forget the name of) not capable of emulation on the MBP yet?
Correct.

Parallels will run the ARM version of Windows, and then you rely on Windows' ARM->x86 emulation layer to run x86 software. VMWare doesn't support it yet, saying Windows doesn't have an official licensing path for their ARM version to support it. But that isn't the same thing as running x86 VMs, which is what all of my VMs currently are.

I don't know if any Hypervisor is exploring using Rosetta to enable running x86 VMs or not, but I suspect not, as then you aren't really running on a CPU, your trying to run on an emulation layer.
 
Tough to tell from pictures, but it looks similar in size to my threadripper.
 
Correct.

Parallels will run the ARM version of Windows, and then you rely on Windows' ARM->x86 emulation layer to run x86 software. VMWare doesn't support it yet, saying Windows doesn't have an official licensing path for their ARM version to support it. But that isn't the same thing as running x86 VMs, which is what all of my VMs currently are.

I don't know if any Hypervisor is exploring using Rosetta to enable running x86 VMs or not, but I suspect not, as then you aren't really running on a CPU, your trying to run on an emulation layer.

Even if you could though, why would you want to?

I mean, sure, emulation will work in a pinch, if you absolutely need to run something, but usually efficiency and performance are killed by any hardware emulation, unless you are trying to emulate something really old.
 
Surprise surprise. I bet peak power consumption is high too.
My 3080 12GB FTW3 can pull 450W, and the 12700K 280W, and that's just with taking the power limiters off, not real overclocking.

And this would still run circles around that for actual content-creation workloads aside from maybe 3D rendering. It'd be slower for gaming of course and probably a few other things, but otherwise?

Apple could make an M1 variant today that'd run typical office productivity stuff in a laptop run a week on battery just by optimizing for battery life.

I'm really wondering if that isn't their next step, and I'm also wondering when they're going to reattack the corporate workstation market.
 
Apple could make an M1 variant today that'd run typical office productivity stuff in a laptop run a week on battery just by optimizing for battery life.

It's very nearly there today if you enable Low Power mode. It has no impact on typical Office applications. Might be able to get about the same on the M1 iPads as well - the battery is a bit smaller, but so is the screen and the hardware is a lot more compact, and Office runs over there as well - just have to deal with the touchscreen instead of a mouse.

The screen shot is my MBP right now - last charged date is accurate. I've only used it for light email and browsing since then, but on Day .. 4?

Screen Shot 2022-03-21 at 8.50.42 PM.png
 
That's the basic idea - scale back on processing cores, especially GPU and AI, maybe some of the transcoding stuff and memory and storage bandwidth and optimize for a lower TDP - and maybe dial back on display quality etc., and you have a US$500 MSRP laptop that's built to the highest QC standards and runs everything most consumers / office workers ever need. For a week, unplugged.

Apple's not really known for catering to lower-budget markets on a regular basis, but with their current architecture they're actually the best suited to do so, simply by leveraging their current R&D and command of component supply.
 
I didn't look up actual dimensions, but I was wondering which was bigger, this or EPYC/Threadripper.
M1 Ultra is around 860mm2, Epyc/TR package comes in at just over 1,000mm2, so it is a bit larger.
 
Here is an interesting article from Phoronix. It's probably the most comprehensive cross-benchmarking I've seen comparing M1 to various AMD/Intel products. Here, he's using a 2020 M1 Mac Mini (4p/4e 8 GPU). This is also the first article I've seen benchmarking Linux running on the M1.


The Firestorm/Icestorm combo of CPU cores look like they roughly slot between Zen2 and Zen3, and around the 10th gen Intel CPU -- on an unoptimized Linux platform at least. Getting the optimizations there, or running the suite under OS X might net it a bit, but I don't think the changes would be game changing.

Here he's strictly looking at CPU performance - there is no GPU driver for M1 Linux as of yet. There's also no accounting for power usage or thermals or overclocking - everything stock. Michael also notes that P-states/boost are currently not working correctly under M1 Linux.
 
The Firestorm/Icestorm combo of CPU cores look like they roughly slot between Zen2 and Zen3, and around the 10th gen Intel CPU -- on an unoptimized Linux platform at least. Getting the optimizations there, or running the suite under OS X might net it a bit, but I don't think the changes would be game changing.
That's actually a pretty decent showing for a completely unoptimized OS. Obviously Apple wants no official part in Linux as it goes entirely against their current direction, but they haven't really done anything to inhibit development either. I expect performance will improve, especially as support for more specific usecases become workable.

Here he's strictly looking at CPU performance - there is no GPU driver for M1 Linux as of yet. There's also no accounting for power usage or thermals or overclocking - everything stock. Michael also notes that P-states/boost are currently not working correctly under M1 Linux.
GPU driver, transcoding driver, AI cores driver... lots of stuff that would be useful in addition to turbo too.

Also, I'm wondering how well a Linux VM would work on an M1. I'm not sure how much use that would be, but it might be a good way to explore adapting or blindly porting Apple's API to Linux.
 
Also, I'm wondering how well a Linux VM would work on an M1.
To be honest I don’t know if M1 has hardware level virtualization support (like the VT-d or AMD-V extensions). You would imagine it would at least for security sandboxing, but Apple doesn’t do a whole lot of first party virtualization - and if they don’t need it they aren’t going to fool with adding it.

I know hypervisors can work without it, and at least one third party hypervisor does exist on M1, I just don’t know if there are various degrades in performance associated with it.

On the plus side - apparently you can virtualize most ARM distros
 
Last edited:
On the plus side - apparently you can virtualize most ARM distros
That's what I'm thinking. Not sure if x86 VMs will ever work, but since most of your development isn't arch-specific - Python, web stacks, backendy stuff - just having a local Linux kernel target, like Microsoft has done with WSL for example.
 
Become a Patron!
Back
Top