The M1 Ultra: Apple’s big leap forward in chip design

Apple introduced what it calls “one last chip” to the M1 family, the M1 Ultra at its Peek Performance event. I took a slightly deeper look at what this means.

Delivering the impossible – it’s a day job

The chip was introduced by Johny Srouji, Apple’s senior vice president of hardware technologies. He described the physical limitations to creating chips larger than the M1 Max, explained that the industry generally chooses to make use of dual chips, but described the efficiency loss to that approach.

That’s why Apple chose to combine two M1 Mac chips in the M1 Ultra. “M1 Ultra is another game-changer for Apple silicon that once again will shock the PC industry,” Srouji said. “By connecting two M1 Max die with our UltraFusion packaging architecture, we’re able to scale Apple silicon to unprecedented new heights.”

He also said the M1 Ultra “completes the family” of Mac processors — at least for now.

Srouji moved on to explain how the company has managed to combine two M1 Max chips into one M1 Ultra. He pointed out a die-to-die interconnect tech that was always on the M1 Max, something Apple has not discussed before. (I seem to recall this was spotted by others at some point, but can’t find a reference.)

What is UltraFusion?

That interconnect uses a proprietary packaging architecture Apple calls UltraFusion. Apple describes this as “way ahead” of the industry and claims it can shuttle data around the chip at 2.5 terabytes per second of low latency, inter-processor bandwidth, which is fast.

To do so, UltraFusion uses a silicon interposer with twice the connection density “of any technology available,” said Srouji.

A silicon interposer is basically an in-package interconnect that bridges the two dies used in M1 Ultra. (I imagine Apple may have used TSMC’s 3DFabric technology to achieve UltraFusion.)

Apple won what appears to be its first related design patent in March 2021, though work began earlier. Its approach isn’t unique; AMD and others have also developed silicon interposer interconnection tech for their PC chips. But Apple’s claim of achieving a superior solution will resonate.

Apple says its tech delivers 4x the bandwidth of competing interposer technologies.

Another important element to Apple’s invention is that developers won’t need to rewrite their code to use this power;  UltraFusion means that, on a system level, the Mac sees the chip as a single processor, not two.

What this means in practice is that applications can access twice the transistors and double the performance, efficiency, and GPU cores — without code tweaks.

Meet the M family

Apple’s processor family now offers four variants equipped as follows.

The M1

The critically-acclaimed M1 processor delivers:

• 16 billion transistors and a 119mm squared-die size.
• 4 performance cores, 12MB L2 Cache.
• 4 efficiency cores ith 4MB L2 cache.
• 8 GPU Cores.
• 16GB DDR4x memory at 68GB/s.

The M1 Pro

The M1 Pro takes this higher, with:

• 33.7 billion transistors on a 240mm squared die.
• 8 performance cores, 24MB L2 Cache.
• 2 efficiency cores with 4MB L2 cache.
• 16 GPU Cores.
• 32GB DDR5 memory at 200GB/s.

The M1 Max

The M1 Max provides:

• 57 billion transistors on a 420mm squared die.
• 8 performance cores, 24MB L2 Cache.
• 2 efficiency cores with 4MB L2 cache.
• 32 GPU Cores.
• 64GB DDR5 memory at 400GB/s.

And the new M1 Ultra

The M1 Ultra brings you:

• 114 billion transistors on a 840mm squared die.
• 16 performance cores, 48MB L2 Cache.
• 4 efficiency cores with 4MB L2 cache.
• 64 GPU Cores.
• Up to 128GB DDR5 memory at 800GB/s.

What it all means

Because M1 Ultra has double the cores, it should manage even the most complex computational tasks. Indeed, given that the M1 Pro already delivers what most Mac users need, this chip will inevitably open the door to new, highly intensive experiences. After all, the performance and power you can exploit in 3D design, video, and audio applications should also benefit gaming.

That the M1 Ultra also offers 64 graphics cores is remarkable.

Together, this combination of technologies means Apple has been able to build a chip that’s 1.9 times as powerful as the latest 12^th generation Intel Core i9-12900K CPU when both run at 60 watts. In other words, you achieve more raw CPU performance at a fraction of the energy costs.

That’s nice for an individual user, but for companies running banks of high-end machines, reducing energy consumption has major benefits to business costs.

When running an application, you can also address a huge 128GB of unified memory. Given pro video production studios run banks of Macs performing computationally intensive operations, this makes for significant budget savings. That they also perform so much faster means investments in these machines becomes little more than “a rounding error” for some pro shops, as CCS analyst Ben Wood noted.

Meanwhile, back in the real world

We can’t yet know what the real-world impact of this performance will be, but the first Geekbench scores show a Mac Studio running an M1 Ultra achieves a single-core score of 1,793 and a multi-core score of 24,055.

In contrast, the highest end Xeon W Mac Pro scores 1,152 and 19,951 in the same tests, while AMD’s heavily-touted Threadripper 3990X is only slightly faster with 25,133 multi-core and just 1,213 single-core scores.

Of course, synthetic scores don’t mean much when it comes to real-world use.

While that’s true, I can say that what I’ve seen of M-series chips so far strongly suggests these tests will translate into real-world application speed gains. Given the move to M1 pretty much doubled Adobe application performance on the Mac, I’ll be very interested to see what Adobe says about M1 Ultra in future.

Apple itself claims the processor will deliver 3.8x the CPU performance, 4.5x the GPU performance, and 3x faster machine learning in comparison to the also outstanding M1 Mac and Intel iMac systems.

[Also read: One year on, developers still love Apple Silicon Macs]

What happens next?

Ending its presentation, Apple said it had one more Mac to introduce — but said this would be for another time.

Many assume this will be the Mac Pro, but we cannot be certain whether the company intends to use the M1 Ultra in that computer, or if it has other plans. I think it likely Apple will further optimize the M1 Ultra design and clock the speed up to exploit the additional thermal management available in the Mac Pro.

Apple’s top silicon engineer did allude to design improvements during his presentation, and there’s recognition across the industry that we’re approaching finite limits in terms of processer technology. That means future improvements will increasingly be around die design, packaging, and architecture – and is likely another reason Apple developed UltraFusion: this forms a route to further chip design innovation.

Unified memory and continued extension of what is placed on the SoC (such as adding Apple’s own 5G modem design) will unleash further evolution, as does inclusion of a GPU on the same processor.

Apple should also be in position to introduce M-series processors built on 3nm design processes once it begins migrating to 3nm manufacturing in 2023-24. This will inevitably enable it to introduce more power, performance, and efficiency enhancements, even as the rest of the industry crawls toward 5nm chips.

What the business thinks

Pointing out that the M1 Ultra is 7x faster than the M1, Wedbush analyst, Daniel Ives, told clients this represents “a staggering technology achievement that is another shot across the board at chip stalwarts such as Intel as Apple beats chip companies at their own game on the desktop front.”

“Apple has already set a rather high bar with the M1 Max, and now they’re aiming to exceed it with the M1 Ultra. And if they can deliver on those goals, then they will have twice set a new high point for SoC design in the span of just 6 months. These are exciting times, indeed,” wrote Anandtech.

However, one big question remains unanswered: why does the Mac Studio with an M1 Ultra weigh 2 pounds more than the one with an M1 Max? The answer, apparently, is that both have the same built-in 370-watt power supply. The additional weight reflects that the M1 Ultra model hosts a larger copper thermal module in comparison to the M1 Max Mac Studio, which uses an aluminium heatsink.

Please follow me on Twitter, or join me in the AppleHolic’s bar & grill and Apple Discussions groups on MeWe.