Ever since Apple bought PA Semi, it’s been clear the company had aspirations of semiconductor design. Its CPU architectures have differed from other vendors by emphasizing raw single-thread performance, and it’s been debuting co-design chips like motion and audio co-processors for the last several years. It’s also purchased part of Toshiba’s flash memory business, thereby gaining a NAND source for itself without having to directly shoulder all of the burden of building a fab and NAND business. And Apple claims its most recent chip, the A11 Bionic, includes a neural engine for AI workloads. The company hasn’t revealed how this chip works yet, so we don’t know if it’s a more conventional, repurposed piece of silicon specifically dedicated to AI tasks, or if Apple is using part of its custom GPU to handle these workloads.
As Nikkei reports, Apple has other significant opportunities to integrate additional functions into its own SoCs. The image below shows the various suppliers of its components. While it’s slightly out-of-date (Apple is no longer using Imagination Technologies as a supplier), it still makes the point. We tend to think of a phone as being comprised of an SoC, some RAM, a camera, and various screen-related components, sometimes with a co-processor to handle motion processing. But there are still major opportunities for Apple to innovate and add functionality to its own SoCs.
Image by Nikkei
By designing its own chips, Apple can better differentiate itself from others, said Mark Li, a Hong Kong-based analyst with Sanford C. Bernstein. “Further, depending too much on other chip suppliers in the age of artificial intelligence will deter its development.”
Apple has been rolling out these integrated capabilities over the last nine years, steadily taking over more of the phone’s logic in the process. Supposedly, Apple has its sights on baseband cellular modems, which would be a huge shift from the current status quo. Both Qualcomm and Intel sell Apple modems for its iPhones, and neither company is going to be excited at the idea of losing that business.
At the same time, however, building modems from scratch is hard. Building modems from scratch that won’t get you sued for patent infringement from other companies is even harder. Thus, while it’s tempting to say that Apple building its own modem would mean we’d see this feature in 2018 or 2019, they’re highly unlikely to be done that quickly. Modem technology is a field where established semiconductor companies often struggle, though whether that’s due to Qualcomm’s business practices or the intrinsic and acknowledged difficulty of designing a modem that won’t be encumbered by patents is a different question.
Given that 4G and LTE are now mature markets, Apple will almost certainly target the 5G standard. Obviously any device that supports 5G will have to have fall-back capabilities to 4G or LTE, just as 4G and LTE devices can fall back to 3G or even EDGE. And of course, the perennial argument is trotted out that Apple wants to build Mac systems that can replace Intel processors. I’ve never said that this is impossible–it isn’t–only that it’s unlikely. And for all the progress Apple has made in CPU design, I think that’s still the case.
Building a high core-count CPU–and I’m defining that as four or more “big” cores–is harder, in many ways, than you might think. Simply slapping core designs down on a floor plan isn’t sufficient. Chips need interconnects, cache allocation strategies for L2 and L3, bus topologies, and power gating. These systems are all highly complex in their own right, and building a unified L3 across 12-28 cores isn’t easy. There’s a reason why AMD’s Zen architecture uses a CCX structure with an 8MB L3 cache allocated per quad-core, as opposed to a completely uniform L3 across all 16-32 Threadripper / EPYC cores.
Could Apple solve these problems? Absolutely, yes. But it’ll take time to ramp up that kind of project, and Apple would have to negotiate a tricky path forward. Every person who bought a Mac laptop so they could run Windows applications on it–and while that may not be a majority of Mac users, it does constitute some of the base–would be left in the cold unless Apple could simultaneously negotiate deals with the developers of that software to port it to macOS. Again, could it? Sure. When you’ve got a few hundred billion in the bank, “persuading” other companies to do things is typically just a matter of cost.
Apple will likely lay these plans quietly and carefully, and take several years to do it. If I had to guess if we’d see an Apple 5G modem or an Apple ARM-based Mac Pro first, I’d bet on the modem–and I wouldn’t expect to see that before late 2019 or 2020.
Now read: How L1 and L2 CPU Caches Work, and Why They’re an Essential Part of Modern Chips