The A5 changed underneath the iPad 2 this spring, and Apple did it without a word. A die-shrunk version slipped into the cheaper iPad 2,4 and the new Apple TV, same performance, a chip about 41 percent smaller and cooler, built on Samsung’s 32nm process. No press release, a year into the patent war Apple is fighting with the same company that builds the chip. Apple’s most important manufacturing partner and its loudest courtroom adversary are still one company, and 32nm A5 wafers roll off a Samsung line while both sides prep for a trial this summer.
32nm is where Samsung makes a real technical bet, and it is not the bet everyone else is making. The node means committing to a high-k metal gate, a new recipe for the part of the transistor that switches it on and off, and there are two ways to build it. You form the metal gate early, gate-first, or you build it near the end, gate-last. Samsung commits to gate-first, which is cheaper and denser when it works and less forgiving when it does not. The pure-play foundry across the strait went the other way, gate-last at 28nm, paying more to dodge exactly the headaches gate-first can cause, while Intel has built its high-k transistor gate-last since its 45nm parts in 2007. Same physics, opposite answers, and Samsung is the odd one out.
The shrunk A5 is a clean read on whether the gamble holds. Same dual-core Cortex-A9, same graphics, identical performance, but the die falls from about 122 square millimeters to 69, and the battery runs an hour longer. Building on the 45nm work that produced the A4, Samsung gives Apple the same chip in far less silicon. Tellingly, Apple put 32nm on its low-volume parts first, the older iPad 2 and the Apple TV, while the headline new iPad this spring keeps its bigger A5X on proven 45nm. That is a company dipping a toe, validating Samsung’s newest line on products it can afford to get wrong before the chip in the next iPhone rides it.
The entanglement only deepens. The same Samsung 32nm gate-first line is building the brain of Samsung’s own new flagship, the quad-core Exynos inside the Galaxy S III it just announced, the phone aimed squarely at the iPhone. Apple’s chip and its rival’s chip, again, off the same Korean process. And all of it runs under a lawsuit that started last April and is now grinding toward a landmark California trial, the two companies spending a fortune fighting over rounded corners and home-screen grids while the purchase orders for A5 wafers go out on schedule, signed by people who will be in a deposition by Friday. Perso, this is the part outsiders never quite believe: the headlines say war, the supply chain says business as usual.
Somewhere in here Samsung stops treating the foundry as spare-capacity rental and starts treating it as a business worth winning on purpose. The gate-first decision is the tell. You do not pick a riskier, denser process recipe to soak up idle time, you pick it because you mean to compete on density and price against the foundry that owns the market. Samsung is signaling it wants to be chosen for leading-edge logic, not merely tolerated for overflow. It is still a distant second to TSMC’s roughly half of the foundry market, and most of its leading-edge volume belongs to a single customer, but the ambition is now out in the open.
The gate-first gamble is the thing to watch. If it yields, Samsung walks into every customer meeting with a cost-and-density story it can wave at fabless designers shopping for a fab. If it stumbles at the next node down, the whole camp betting on gate-first has to rethink, and the company that went gate-last gets to say it warned them. I think gate-first survives at 32nm and gets meaner after that, though I have been wrong about transistor recipes before. What is not in doubt is the customer relationship, which looks unbreakable right now, lawsuits and all, and the chip in the next iPhone will almost certainly ride this same 32nm line. Unbreakable things tend to break on a schedule nobody saw coming.