Apple M1 Ultra: explanations around an innovative chip fusion

Apple continues to develop its series of M1 chips and, after having already hit the nail on the head with the M1 Pro and M1 Max, is waving in the face of creators and other professionals concerned with having high computing power with a high-color and in transistors. To sum up the idea behind this fourth version of the M1 chip, it is “simply” the merging of two M1 Max chips into a single chip with 114 billion transistors. Simple to say, but difficult to achieve, this merger hides above all quite innovative details.

Apple M1 Ultra: two M1 Max glued side by side

Before we get to the nitty-gritty, let’s go over everything that’s not changing. The architecture first of all, since Apple purely and simply takes over the die – that is to say the small piece of silicon comprising the transistors – of the M1 Max, without any technical modification. The engraving, then, is always carried out in 5 nm in the TSMC factories, still giving Apple a considerable technological lead over Intel – confined to 10 nm for the moment – ​​and AMD, whose professional Ryzen processors are engraved in 7 nm. The processor and graphics cores, finally, are de facto technically similar to those of the M1 Max, but doubled in number.

On the front, the M1 Ultra chip is presented as a SoC offering 20 CPU cores – 16 high-performance Firestorm cores and 4 low-power Icestorm cores – and 64 GPU cores for 8192 computing units. A wiser version is also offered with the same number of processor cores, but a GPU made up of 48 cores, or 6144 computing units.

At the output of the M1 Max chip, some observers had spotted a small area of ​​transistors taking the appearance of a strip at the top of the SoC. What then let suppose that it was about a means of connecting two dies together. This is indeed the case, Apple using this area to communicate together the two parties, a technology that the firm calls UltraFusion. This technology, made up of more than 10,000 links according to Apple, allows the two dies and benefit for this from a maximum bandwidth of 2.5 TB / s. Enough to allow data to pass in one direction at 1.25 TB / s and as much in the other direction.

You guys seeing this or am I just crazy? The actual M1 Max die has an entire hidden section on the bottom which was not shown at all in Apple’s official renders of the M1 Max die. Just flip another M1 Max and connect it for an M1 Max Duo chip. Then use I/O die for M1 Max Quadra. https://t.co/McWmofJAlspic.twitter.com/JogRwUGvF6

—Vadim Yuryev (@VadimYuryev) December 2, 2021

This fusion of the two dies is then installed on the chip itself (we speak of interposing) and surrounded by LPDDR5 type RAM chips. Apple offers 64 or 128 GB of RAM operating on 32 channels for a fairly high bandwidth of 800 GB / s – four times more than on AMD Threadripper processors, for example, but a little less than on the GeForce RTX 3090 from Nvidia (936 GB/s).

Two graphics circuits that are one

But beyond these technical aspects of assembling two dies, Apple manages to innovate in a completely different way. To explain this, it should first of all be remembered that the company is aiming with its M1 Ultra at an audience of professionals with high needs in terms of hardware acceleration. These are creators using video editing, 3D or post-production software of all kinds. For this type of activity, it is necessary to use workstations with a fast processor, but also and above all equipped with a strong graphic computing power – the GPUs allowing to brutally accelerate the calculation times.

To go even faster, some workstations use multiple graphics card configurations, whether multi-GPU or multi-card. The problem so far encountered has always been to get the different GPUs to communicate quickly enough so as not to end up with restricted performance. The adopted approach will then have been to present to the system (and to the software) these graphics cards as several elements which work in parallel on several tasks distributed less skilfully by the applications.

And that’s where the Apple M1 Ultra changes the game. Although there are physically two graphics circuits, it is a single GPU that is presented to the system and software. It is totally transparent in order to optimize the performance gain (we are talking about scaling). Assuming that the bandwidth is sufficient and that there is no latency limitation, we can then hope to find ourselves in the presence of an easily optimizable solution for software publishers.

Performance at the level of the best consumer chips from AMD, Intel and Nvidia

And performance in all of this? The graphics part is announced by Apple as being about as fast as the GeForce RTX 3090. Rather impressive for a graphics circuit housed alongside a processor. But, even more interesting, this duo of graphics circuits would only need about 100 W to be powered against more than 300 W for Nvidia graphics cards.

The processor would be around 25% faster than Intel’s Core i9-12900K (16 cores). Again, the power footprint would be heavily in Apple’s favour, with the processor part pulling around 60W versus Intel’s 160W. Amazing data knowing that the i9-12900K actually pulls 240W.

Still, based on our tests conducted on both the latest desktop chips from AMD and Intel, as well as on Apple’s mobile M1 Max, we can already paint a sketch of the expected results. In “processor” video encoding under HandBrake, Apple’s chip should be at the level of a Ryzen 9 5900X and a Core i7-12700K. In “processor” 3D rendering, these two competitors could be overtaken without the M1 Ultra being able to reach the higher benchmarks. The unknown lies rather in the ability to produce results when applications allow the use of the onboard graphics card to speed up processing. In this case, the M1 Max generally offering a service close to a desktop GeForce RTX 3070, it is indeed not excluded that the GeForce RTX 3090 is equaled.

The Mac Studio, the first product to integrate the M1 Ultra

To conclude this first approach to the M1 Ultra chip, we can only rejoice that Apple is bringing a technological breath of fresh air in the field of hardware acceleration. The multi-designdie of this chip is innovative and the energy efficiency seems to be far ahead of anything offered by the competition. In the end, you only have to look at the size of the Mac Studio to be convinced: who could still have imagined that a workstation at least as fast as the combination of the best competing processors and graphics cards can fit into such a small package?

Still, it is for the moment a promise and the field test will be much more revealing of the capabilities of this chip. We look forward to offering you a hands-on experience with the Mac Studio so that you can judge the behavior of its M1 Ultra chip in professional applications and measure its energy impact when plugged in, but also in terms of noise pollution.