Review: Apple Mac Studio with M2 Ultra SoC

Like a car enthusiast that gets excited by the latest fast sports car, I get excited about fast computers. I have absolutely no need for a fast computer as most of my tasks are mundane and run fine on a 5 year old computer, just like I’m never going to take my car to a race track, but I still enjoy seeing the benchmarks on the latest SoCs/CPUs and going “oh wow that’s quick” when the numbers are a few percent bigger than the same time last year.

Because Apple have a 14-day no-questions-asked return policy and I thought it would be something interesting to write about, I abused my credit card to get Mac Studio featuring the brand spanking new M2 Ultra SoC.

If you’re unfamiliar with the M2 Ultra, it uses TSMC’s N5P process, a refinement of the N5 process used in the Zen 4 CPUs (Ryzen & EPYC) made by TSMC for AMD. It’s essentially the most advanced semiconductor tech on the market. The smallest of transistors money can buy! The M2 Ultra has 24x ARM64 cores, 16 of which are “high performance” and 8 are “high efficiency”.

As Apple explains, the M2 Ultra is two M2 Max “dies” stitched together through a 2.5TB/s interprocessor interconnect, but appears as a single CPU to the system. Unlike traditional CPUs, the M2 Ultra also includes memory integrated into the SoC, enabling a whopping 800GB/s of memory bandwidth. DDR5-5200 memory in a “traditional” PC has a maximum of around 64GB/s. There’s also a decent GPU built-in, with 60-cores.

The only high performance thing I do regularly is crunching a shitload of TIFF files into RAR files for uploading to the Internet Archive (which only supports RAR & ZIP unfortunately), so that’s where I started.

My main rig for that task is a Lenovo P700 workstation with 2x Intel E5-2670 v3 (24C/48T) CPUs and 128GB of RAM with a GTX1650 GPU running Ubuntu 23.04:

time tar --use-compress-program="pigz -9 -k -p48" -cvf /media/decryption/samsung250/out2.gz ~/apc/1980_01/*.tiff
1m 5.885s

time ~/rar/rar a -ep -ep1 -htb -ma5 -md128m -rr10 -mt48 -s /media/decryption/samsung250/out2.rar ~/apc/1980_01/*.tiff
6m 6.981s

The Apple M2 Ultra (24C/24T) did the same thing (running macOS Ventura):

time tar --use-compress-program="pigz -9 -k -p24" -cvf /Volumes/sam1tb/apcziptest2.gz ~/Documents/apc/*.tiff
1m 0.22s

time ~/rar/rar a -ep -ep1 -htb -ma5 -md128m -rr10 -mt24 -s /Volumes/sam1tb/out2.rar ~/Documents/apc/*.tiff
3m 32.83s

Not surprised that the rar command completes in half the time, but I am surprised the pigz command (parallel zip) is only 8% faster on the M2 Ultra. Unlike WinRAR, the CPU was going at full tilt on both computers, so it’s not like the disks were holding things back.

The upside however is that the M2 Ultra did it with using less than half the power. Roughly 80W with the pigz test with CPU at 100% compared to over 200W on the dual Xeons. It’s also virtually silent and puts way less heat into the room.

I know it’s unfair to compare a CPU from 2015 to the M2 Ultra, but it’s all I’ve got on me at the moment! Nobody sells PCs with an as generous return policy as Apple so it’s hard to get my hands on a modern Ryzen 7000 or 13th-gen Intel PC.

Got any tests you’d like me to run on the M2 Ultra for you? Send me an email (aagius@fastmail.fm) or join The Sizzle’s Slack channel and look for @decryption

7zip benchmark & ARM/x86 virtualisation

Windows has a native ARM version that runs quite nicely in a virtual machine on Apple Silicon Macs. Inside that ARM version of Windows is a translator that lets you run x86 apps, with a performance penalty of course.

A friend on Twitter asked me how the M2 Ultra goes running x86 Windows apps in a virtual machine because he still runs some CPU heavy x86 Windows apps on his Intel Mac, but knows eventually he’ll have to upgrade to Apple Silicon.

Here’s the 7zip benchmark running natively in macOS on the M2 Ultra:

Here it is running in the ARM version of Windows 11 in a VM with Parallels, using the ARM version of 7zip, configured with 32GB of RAM & 22 cores (because Windows doesn’t like odd numbers and giving all 24 cores to the VM isn’t fun for the host OS):

Here’s the benchmark running in the same ARM Windows 11 VM, but the x86-64 version of 7zip (i.e: using the built-in Windows ARM to x86 emulation/translation):

And finally, here’s the same benchmark on the 2x Intel E5-2670 v3 (24C/48T) CPUs rig on Ubuntu:

The key figure is average Rating MIPS

  • M2 Ultra macOS native: 183,496
  • M2 Ultra Win 11 ARM native (22 cores): 144,162
  • M2 Ultra Win 11 x86 emulated (22 cores): 97,501
  • 2x Intel E5-2670 v3 Linux: 94,573

The x86 emulation is much slower than native, but still faster than my dual Xeon rig! That’s bloody impressive

You can see a bunch of other results for the same test on Open Benchmarking

183,000 MIPS puts the M2 Ultra in the same league as the Ryzen 9 7950X and the i9-13900K - much cheaper computers than the Mac Studio. You can build a nice Ryzen 7950X rig for ~$4k, more on that in a separate post.

Running in a VM under x86 emulation in Windows, the M2 Ultra performs like an Ryzen 9 3900X, which feels suspicious to me, but this benchmark scales well based on the number of cores and we’ve got 22 cores here compared to the Ryzen 3900X’s 12 cores. So while the emulation might be slower, there’s almost twice as many cores to make up for it.

I gotta do more testing on this I reckon. Feels almost too good to be true when running x86 apps in a Windows 11 VM.

Chucked macOS Sonoma on the Mac Studio and installed the Game Porting Toolkit as per the instructions inside the GPT (that acronym won’t be confusing!) package and the Apple Gaming Wiki.

The only fancy modern game I have in my Steam library is GTA5 and it’s not really that modern. I gave it a crack anyways. Here’s a benchmark run at 1080p with the graphics settings all turned up to the maximum:

This is with no tweaking at all and it achieves a relatively stable 30FPS or higher until the very end. Considering the entire thing is being translated from x86/DirectX to ARM/Metal, it’s hard to complain.

I’m gonna try get some more games and do some benchmarks. If anyone wants to send me some games on Steam, my username is decryption256 :slight_smile:

I tried to install Ashai Linux today, but didn’t realise it’s limited to M1 Macs only and not even the Mac Studio M1, let alone M2. So any Linux tests won’ t be happening now, sorry :frowning:

After having spent a few days with the M2 Ultra, it’s an impressive little beast. It runs so quiet and uses fuck all power.

The max power draw I ever saw it reach was around 250W when the GPU and CPU were being used at 100%. With the CPUs alone the max it drew was around ~120W.

There’s nothing on the market so efficient and available off the shelf. When purchased on sale at JB Hi-Fi during a 10% off sale, with 10% off gift cards, you can get it for $5,345.19! Sounds like a lot of money, but for a similarly powerful PC, you’d be looking at around $6200 anyways.

The 7950X3D performs around the same as the M2 Ultra depending on the task and the Nvidia RTX4090 is way more powerful. The downside is you’ve got to build it yourself (not an issue for some people, but for others it’s a challenge), it’ll use a heap more electricity and is louder, hotter and physically larger.

You’ve also got your consider your workflow. If what you’re crunching away on is optimised for x86, well the M2 Ultra isn’t going to help you. If you need CUDA instead of Metal for your GPU work, the M2 Ultra is no good. Need Windows or Linux? The the M2 Ultra is useless. But if your workflows can be adapted to ARM and Metal, the Mac Studio is an awesome machine and actually excellent value for money.