Article

Apple Silicon: Ultra Clever

· Ben Bajarin

I have been arguing for years, and I will die on this hill. Apple silicon remains one of the most innovative products Apple creates on an annual basis. You may disagree with the framing of Apple silicon as a product, but knowing Apple as I do, they view almost everything they create as a product even if it is an ingredient like Apple silicon. Saying Apple silicon is innovative or the most innovative does not take away from product innovations that come to iPhone, iPad, Mac, AirPods, etc. Still, it recognizes the amazing science that goes into silicon design and how difficult it is to truly innovate around silicon.  Viewing Apple silicon as a product also helps us better understand how it is a differentiator for Apple for every product they release with Apple silicon inside. To quote Tim Cook, “Apple silicon helps us create products we could not imagine before.” For example, during the Mac part of the presentation, Apple’s speakers went out of their way to emphasize how Apple silicon has transformed the Mac and insinuated it was a contributing factor to Mac’s record-breaking quarters and growth outpacing the PC industry. I have long believed that Apple could not truly build the Macs they wanted until they put their silicon inside, and I am more convinced than ever that is true.  Before I dive into the M1 Ultra and its significance and potential foreshadowing of future Apple processors, I want to emphasize what Apple’s goal, architecturally, is with their silicon strategy. It should be no secret because executives from Tim Cook to Johny Srouji repeat this phrase that Apple’s design ambition is to lead the industry in performance per watt. Other companies may create laptops or workstation SoCs with more performance than Apple’s, but Apple’s goal is to beat them in performance per watt. Competitors may come up with SoCs that have longer battery lives than Macs, but they won’t have the performance and battery life of a Mac. We can not fully analyze Apple silicon and its role in Macs without always bearing in mind Apple’s goal of leading in performance per watt.  Enter the Ultra
 Apple’s commentary during the event was their newest chip is the last chip in the M-Series family which now consists of M1, M1 Pro, M1 Max, and M1 Ultra. The M1Ultra is the newest edition and perhaps the most clever design that also does some future signaling of Apple’s architecture evolution. With the family now set, we can expect the next M-series family to be M2, M2 Pro, M2 Max, M2 Ultra, and so likely an annual cadence. But the M1 Ultra does something extremely clever that gives Apple a number of interesting advantages architecturally.  M1 Ultra is, to my knowledge, the first die-to-die or monolithic SoC to Monolithic SoC implementation. Apple calls the architecture they used to fuse these two chips together UltraFusion. Apple used a technique using an interposer at the substrate level as well as the silicon package itself to allow for 10,000 connection points, thus rendering two M1 Max chips as one single monolithic SoC. This approach allows the software to see only one SoC but leverage all the CPU cores and GPU cores as if it was a single chip rather than two.  Remarkably, this approach is potentially more power-efficient than if Apple simply designed a new SoC the die size of the M1 Ultra. This technique also offsets many of the challenges in developing a single monolithic SoC the same size as M1 Ultra. Apple’s own charts show that M1 Ultra is still extremely efficient from a performance per watt standpoint. Fusing these SoCs together leverages the efficiency of the M1 Max design and combined gets double the performance with approx double the power consumption. Again, it underscores Apple’s efforts to be the industry leader in performance per watt. Where this gets really interesting to me is how Apple is starting to innovate at the package level of silicon, not just the architectural level of design. This is extremely relevant because as leading-edge process nodes slow or become harder to achieve on a predictable cadence, the packaging will start to matter more to get additional gains and efficiency on all performance vectors per watt.   This approach helps Apple on several fronts. It gives them a solution for when semiconductor foundries hit a wall with process technology. While it seems clear, 3nm is achievable, there are still question marks on the most efficient way to get to 2nm and even more questions for 1nm and beyond. There is a real chance the predictable cadence of process technology Apple has seen from TSMC slows significantly after 3nm, and nearly inevitable that getting beyond 3nm continues to increase in costs.  By having an approach, as Apple has taken with M1 Ultra, Apple has come up with a way to hedge against process node advancement slowdowns/challenges and the rising costs of leading-edge processes. While I can’t confidently say Apple will take this approach with future SoCs in the family beyond the Ultra, it certainly gives them the option to apply this approach to future M Pro or M Max SoCs and have longevity both at a current mature process node as well as the economics of mature nodes.  Foreshadowing and Competition
While we are on future speculation, having this design and package innovation on hand would stop Apple from leveraging this die-to-die design and connecting more than two SoCs? For example, could M2 Ultra have four M2 Max’s connected via UltraFusion? And suppose this architecture design does continue to be highly power efficient. In that case, it could make more sense to leverage this approach frequently and reap the rewards of the dramatic increase in performance per watt it delivers.  So what does this mean for the competition? Well, it presents a challenge. Nearly every academic article on the use of silicon interposers points out that this approach favors vertically integrated players. This does not rule out someone like Intel, AMD, or Qualcomm to take this Approach if needed but they are likely to hit a wall at the end-product level. Apple’s advantage is that they design all the core silicon bits but also make the products that silicon goes into. Apple’s verification process is quite different from silicon competitors’ verification processes who have to support a much wider range of products. Intel, AMD, or Qualcomm can take this approach, but the same advantages Apple gets as a much more fully integrated player will not all extend to others in the same ways. Where it makes sense for Apple to take this approach, it may not make the same sense for their competition.  Not too long ago, people speculated the Mac was dead, or a forgotten product. People ranted that Apple has abandoned creative and studio professionals. Looking back, it seems Apple was just waiting on Apple silicon to fully execute their vision for the Mac and build the laptop and desktop computers worthy of their customers.