M1, M1 Pro, M1 Max, and now M1 Ultra. The family of the first generation of Apple Silicon chips for Mac is complete, and this is not anecdotal, for several reasons.
Because it shows that Apple’s acute control over hardware is crucial and gives it great freedom, materialized by a new range of Macs, which hasn’t happened for ages. Unless we’re mistaken, we have to go back to the MacBook Air in 2008 to find a new form factor and a new positioning for Macs.
Because that means the next step will obviously be an M2, with all the surprises and power you can imagine.
But also and above all because we now have before our eyes the cards that Apple intends to play both on laptops and on real workstations. After the arrival of the MacBook Pro at the end of last year, we wondered how the company was going to be able to surpass the M1 Max, an overpowered chip, with an impressive performance per Watt ratio, with its 57 billion transistors engraved in 5 nm, and of an already colossal size with a surface of 423 mm².
The M1 Ultra is the answer to this expectation and to these questions. It is reserved for the moment for the brand new Mac Studio, but could well also take place in a Mac mini “Pro”, if it still makes sense.
A bit of industrial genius
But, even more interestingly, it’s Apple’s approach that captures the attention to achieve the monster that is the M1 Ultra. The American giant plays on the very advanced scalability of ARM chips, with an increase in the number of cores, but with a little technological magic trick of its own. A trick that was planned from the design of the M1 Max, and which had obviously been kept secret.
Rather than create a huge new chip, Apple has chosen to juxtapose and merge two SoCs into one, and this is a first in the history of semiconductors.
From an industrial point of view, it’s a stroke of genius – and an answer to an old fantasy. The M1 Max had reached a good number of design limits, in terms of size, in particular. However, more one die is large, the more difficult it is to produce and the lower its rate of return. Not only would designing a larger chip require heavy investment, but manufacturing it would have been complicated. The teams of Johny Srouji, Apple’s vice president of hardware technologies, have therefore opted for another path. In this case, Apple takes advantage of what it has mastered and experienced.
This allows him to double just about everything. Starting with the number of transistors, which increases to… 114 billion. The M1 Ultra steals its crown from the Ponte Vecchio graphics processor, from Intel, and its 100 billion transistors. And of course, that doubles everything. The number of CPU cores, GPU cores, neural network cores, first. But also the controlled inputs/outputs for the peripherals, the memory bandwidth (800 GB/s!), the amount of LPDDDR 5 RAM managed, now up to 128 GB, etc. RAM is still unified, i.e. shared between CPU and GPU cores, and its modules are soldered directly to the SoC.
Doing something new with “two old people”
Apple therefore pulls off a little magic trick, juxtaposing two dies on a single package, two M1 Max, to make one M1 Ultra. The various representatives of the Cupertino giant have insisted on this point: the Mac and its software do not perceive its two dies merged only as one and the same chip. A single CPU, and especially a single GPU – which is very promising for multi-threaded tasks, obviously, but a little less so for single-threaded tasks.
This approach is in tune with the times, however, since chiplets are on the rise. To the point that AMD, ARM, Intel, Qualcomm, Samsung and TSMC have recently created a new standard to define how dies of silicon can be combined in a single package.
UltraFusion: the M1 Ultra fulgure, to the point?
The M1 Ultra therefore aligns twice 32 cores, i.e. 64 GPU cores, and 20 CPU cores, divided into 16 high-performance Firestorm cores, and only four low-power Icestorm cores – a Mac connected to the sector does not necessarily need to be careful. to his energy appetite.
Always as hard to find names that say things, Apple calls this encapsulation architecture UltraFusion. To make it possible, Johny Srouji’s engineers used a silicon interposer that connects the two dies of M1 Max from below via more than 10000 points, and routes the signals between the two chips.
It is this interposer who is in charge of communication. By reducing connection latency and offering 2.5 TB/s bandwidth, it makes two M1 Max equal…one M1 Ultra.
This “internal” bandwidth on the chip will allow colossal data movements between the two GPUs, even if this does not mean that the performance is necessarily doubled.
This merger is in any case completely transparent for program developers, which means that it will not be necessary to optimize the code to take advantage of this increase in power, as has been the case in the past for dual-processor graphics cards, for example.
This technology also allows Apple to be very ambitious and confident. The M1 Ultra would be able to offer the same level of relative performance as an Nvidia GeForce RTX 3090 (accompanied by a Core i9-12900KF, from Intel) while consuming 200 Watts less, and just over 100 Watts .
If it will obviously be necessary to see what this gives with tests, we note here the advantage of the ARM architecture and its variation in Apple Silicon. This is the strength of the multiplication of GPU cores, and the impressive number of on-board transistors. Of course, the 114 billion transistors of the M1 Ultra are not dedicated to the graphics part. However, the chip that drives the RTX 3090 only embeds 28.3 billion of them, engraved in 8 nm, by Samsung, and not in 5 nm by TSMC. However, the Taiwanese component production genius currently seems to have a definite advantage over its Korean competitor in terms of manufacturing.
With so many cores, Apple can avoid overstretching them, which means that the operating frequency is not too high and the resulting power consumption is contained.
A dreamy future
Obviously, Apple does not want to say if in the future we will be entitled to a Super UltraFusion, with even more dies merged. However, with this first step, the Cupertino giant will not only continue the technological trend of increasing the number of cores on the CPU side, but also initiate a new movement on the GPU side.
The “fused” multiGPU is a promising, but precarious balance, requiring tight control over hardware and software. However, Apple has control over these elements. If its promise is kept, it could well hurt the competition very badly.
Not necessarily by raw power, but in his own way. Apple is not, like Qualcomm, Intel or Nvidia, a company that manufactures chips that will fit in a wide variety of devices. It is a giant that designs custom SoCs not only for specific uses, but also for specific machines. A luxury as much as a potential weakness in some respects.
Right now, the M1 Ultra frolics in the all-new Mac Studio, there it serves one vision, one product. However, we can’t help but already look to what will happen later this year. If the Mac Studio is an in-between, between Mac mini and Mac Pro, what monster will Apple design for its most powerful computer? Right now, the newcomer is killing the Mac Pro powered by Intel and AMD… That promises some more exciting announcements.
The Apple Silicon revolution is underway. The rise of ARM carries it, as much as it is carried by the M1s. Already, Qualcomm promises to enter the dance more willingly, with its own approach, its own issues. Intel and AMD are obviously not going to sit idly by. The PC world was thought to be dead or, worse, boring. But here, no risk of getting bored, and what a sight!