Sunday, March 20th 2022
Apple Mac Studio Taken Apart, Reveals Giant M1 Ultra SoC
Max Tech performed the first detailed teardown of the Apple Mac Studio, the most powerful Mac since Apple dumped Intel for processors in favor of its own silicon based around high-performance Arm chips built from the ground-up for its own software ecosystem. The M1 Ultra SoC powering the Mac Studio is its most striking piece of technology, with Apple attaching some very tall performance claims not just for its CPU compute performance, but also graphics rendering performance.
The M1 Ultra SoC is physically huge, with roughly similar package size to an AMD EPYC processor in the SP3 package. An integrated heatspreader (IHS) covers almost the entire topside of the package. Things get interesting under the hood. The M1 Ultra is a multi-chip module of two M1 Max dies connected on package using Apple UltraFusion, a coherent fabric interconnect that allows the various components of the two M1 Max dies to access memory controlled by the other die. Speaking of memory, The M1 Ultra features up to 128 GB of LPDDR5 memory that's on-package, This memory is used for the CPU, GPU, as well as the neural processor, and has a combined memory bandwidth of 800 GB/s. The M1 Ultra features up to 20 CPU cores, up to 32 Neural cores, and up to 64 GPU cores (8,192 programmable shaders).
Source:
Wccftech
The M1 Ultra SoC is physically huge, with roughly similar package size to an AMD EPYC processor in the SP3 package. An integrated heatspreader (IHS) covers almost the entire topside of the package. Things get interesting under the hood. The M1 Ultra is a multi-chip module of two M1 Max dies connected on package using Apple UltraFusion, a coherent fabric interconnect that allows the various components of the two M1 Max dies to access memory controlled by the other die. Speaking of memory, The M1 Ultra features up to 128 GB of LPDDR5 memory that's on-package, This memory is used for the CPU, GPU, as well as the neural processor, and has a combined memory bandwidth of 800 GB/s. The M1 Ultra features up to 20 CPU cores, up to 32 Neural cores, and up to 64 GPU cores (8,192 programmable shaders).
41 Comments on Apple Mac Studio Taken Apart, Reveals Giant M1 Ultra SoC
Boy where have we seen this before... ARM? :) Its yet another iteration but I don't see Apple's path forward as something we can use outside of their ecosystem. The package size has no business in any normal competitive market, Apple is the outlier and you can have only one of those. They carved out their own segment, and there is only a place for one Apple in that basket.
Its nice to see what they're doing though, power to them. But I don't view it as a threat to any other market really. Maybe we will move to bigger package in due time. Maybe not. Time will tell. But its not disruptive in any way of other progress like the push on chiplets and/or more specialized cores/big little concepts.
The only real move forward if we zoom out is that device capabilities are moving closer together regardless of device form factors. Phones and small devices become more powerful, all devices can utilize cloud processing power, etc. But the overall demands don't really change, even with the introduction of cloud to local productivity. The real question that will eventually surface (and already is, if you are aware of it) is how much control you want to have over your hardware/software solutions, how much control you need, and what price you're paying for it. Apple isn't in the best position in that sense for significant user/target markets, no matter what chips they choose to build.
The differentiation isn't really in hardware anymore but in policy.
I like that consumer power at work, a lot. I hope MS will keep listening, because its for their own good...
Linux is a real contender sooner rather than later, too, because a big part of Windows in consumer markets right now is gaming. And that really is the three flavors we have:
- open, subject to change you can choose to implement, full adaptability, low user friendly score (Linux)
- not open, subject to change you must implement, high adaptability, high user friendly score (Windows)
- walled garden, subject to change you won't like and will cost you money you don't prefer spending, ultra high user friendly score, very limited adaptability (anything Apple)
With regards to control, PS/Xbox/Switch are some of the most controlled devices out there and yet every one of them is successful. Most people don't care too much about control, they care about experience and walled gardens is the easiest way to manage experience.
People started using tablets and smartphones and you might say 'they can game on it too!' and yet, still, PC gaming shows growth YoY. That's a lot bigger than we are trained to think based on the marketing and 'smartphone centric' world we appear to have, where apparently there is only a demand for the easier and more service-oriented society. This is not true at all and we have yet to see it stick.
Look at the rapidly changing sentiment wrt cloud gaming and on-demand. As the number of (potential) subscriptions grows, so too does the resistance. And these services are all still pushing hard on introduction price structures, freebies, 1 dollar subs and big marketing budgets.
The package of the APU is unconventional but if that level of integration is key to the performance they are getting then either AMD and Intel will have to do something similar or someone else will eventually do it for them. I mean if you use any of those performant applications (video editing, 3D rendering) and if Apple starts walking away from the competition you will switch so in that sense they are absolutely a threat. Users of professional software packages are not allegiant to platforms and/or hardware fanboi/gurls.
The way forward for Apple:
- bigger dies / more chiplets = further cost increase per chip
- even further integration which implies even further tailor made solutions, which kills expandability, unless they can automate it somehow. Which could be their higher purpose, they already built pretty strong recompiling software if I recall to bridge ARM > x86.
Both things are finite. You can only scale chips as far as package allows, and we're already looking at a massive package compared to the competition. You can also only scale chips as far as is economically feasible, and that's a moving target, but still finite in some way.
The way forward for x86:
- the focus is still on small dies, the monolithic die is yesterday's news, so we're looking at not just die size increases but simply more chiplets or better arranged core complexes, but still with a focus on reduced yield risk, ergo, small dies. Nodes get smaller, so the gain here is massive - simply because its achievable. Same die size on a smaller node is already a larger floor plan.
- the focus is on more specialized cores that - again - take up lower square mm per core on the die.
- software efficiency is a per-case scenario. Some software will optimize for newer hardware, other stuff will lag behind, but eventually, economics dictate you will need to optimize to keep up. These are costs Intel and AMD are not making, while Apple forced itself into that software garden, with control comes a large responsibility there.
As Apple dies go larger, I predict they'll face an ever more difficult economic balance with a large package. Part of that can be justified by its performance, perf/watt, and the performance of applications that run on it. But the hard cap of its capabilities will be lower than what x86 can be stretched towards, or Apple will have to sacrifice power efficiency for clocks. Basically... something's gonna give one way or another.
Huge chips and huge packages are not risk free, the real, core question is whether Apple timed their move to a larger package right, really.
So far, historically, every company that could hold on to a smaller chip at a competitive level longer than the rest, is the company that won that specific round of silicon wars. Intel during their quad core / single thread focused days (which is the reason they're not abandoning that race either even with the newest core designs; the reason E cores exist is so they can push P cores harder within similar TDPs), and Nvidia ever since Kepler, and even now, with a feature advantage for Nvidia, AMD manages to strike back with a smaller chip even in the strange market of today. I mean yes, AMD won a few rounds in the GCN years especially with HD7970's, but let's not speak of their margins; meanwhile, Nvidia could keep up with smaller, more efficient designs and they started swimming in gold year over year, even while keeping the performance crown and increasing their lead as AMD's GCN stopped scaling proper past Hawaii XT. It is thát money that enabled them to fortify their lead. We have yet to see how things develop post-Ampere as Nvidia does a proper shrink at last, and not this crappy Samsung business; but the only reason Nvidia had margins to speak of on Samsung is because likely Samsung loved having that high profile business and offered something cheap. Even so... the ride was bumpy and we know yields aren't fantastic, and we also know of several price hikes between Pascal and Ampere, while nearly all of Turing was too expensive even in base MSRP... because the die got huge. One gen post Turing and Nvidia lost a convincing +-3~5-year lead to RDNA2 that now has a far stronger road ahead of itself in terms of die size/scalability.
Also... this is Apple, which kind of lives in a vacuum in tech land and is really happy in it. The reason they are what they are is because they have nice things that aren't for everyone's wallet. It remains to be seen how hard they even want to push and fire on all cilinders. They can easily make do on marketing and minimal progress, as we've seen, again, historically.
I was merely commenting on the cost. The interposer is expensive for that, the motherboard to handle that is going to be some insanely-expensive 12-layer madness...
It likely costs Apple more to make an M1 Ultra than it costs AMD to make an EPYC 7713 which has a $5K list price, not that you can buy one by itself these days.
Again, just because we don't see their roadmap, doesn't mean they don't have one.
I also assume here that they are working closely with ARM to make architectural advancements which will favor their long term objectives. The interesting thing here is that this won't just help Apple in the long run but everyone who uses ARM architecture.
It will be interesting to see what they do for the next generation for their desktop designs and if they change their approach. The M1 at its core is really just stuff they have been doing for iPhones and iPads scaled way up. That totally makes sense as it gives them a very much a good enough starting point but they might have something that really starts to deviate from what they are putting in their mobile devices in ways other than just size for future desktop CPUs.
What they will never achieve is beat PC gaming and raw compute power of desktops, but, the problem is that huge raw compute power desktop's can achieve is most of the time sitting idle in content creation apps and that's why Apple will eventually be overwhelmingly the favorite for some professionals, they put that hardware to work.
Your average windows gamer just wants their game to run, discord or whatever, and some web browser stuff. User-friendly Linux distros that do this and run Proton are getting more popular by the day....