The big news today (it’s always a bit dry after the New Year and before CES) is that TSMC is set to begin trial production of Apple’s A6X processor while ramping A7 production later this year on 20nm. This is a move that’s been a long time coming.
In relationship terms, Apple and Samsung are long past Judge Judy and Jerry Springer and have descended into what sociologists refer to as “The Lohan Zone.” When Apple launched the iPhone 5 earlier this year, it let slip that, while Samsung was building the processor, it was serving strictly in a foundry capacity, having contributed no other IP to the project. That was highly unusual in a day and age when foundries are scrambling to add IDM-like capabilities, and it indicated just how poor relations were between the two companies.
The surprising factor here isn’t that Apple would be moving away from Samsung, but that it’s already taken this long to do so. What that implies is that there are contracts in place between the two companies that have slowed the process. Up to this point, Apple has accounted for the majority of Samsung’s foundry business and it’ll take the South Korean manufacturer time to shift to other customers and keep its fabs running at capacity.
Trial production, meanwhile, is scarcely the same thing as running full out. TSMC has pre-existing contracts it has to honour with Qualcomm, Nvidia, AMD, and the other companies that build on its 28nm process. That’s part of the reason why CEO Morris Chang floated the idea of investing in specific foundry facilities for particular customers earlier this year, and the company has stated it expects to be capacity-constrained on 28nm through 2013.
Nonetheless, getting production of the A6X off the ground is an important step towards breaking away from Samsung. The timing suggests that Apple may have funded the development of two different versions of the same chip – one at Samsung, built using gate-first technology, and one at TSMC using that company’s gate-last process.
It’s entirely possible that TSMC will ramp a 28nm version of an Apple chip up in 2013; the ramp and subsequent demand could even explain why the company has stated it believes 28nm capacity will be constrained through the end of the year. Either way, Apple’s decision to shift to TSMC, when it eventually happens, could have a profound impact on the rest of the industry. If TSMC and Apple begin working closely together, Apple’s competitors may decide to shift their business elsewhere, particularly if the Taiwanese foundry actually built a custom fab for its Cupertino customer. Nvidia has already done some collaboration with Samsung; other players could easily follow suit.
This is a situation to watch. With process node jumps becoming more difficult, both foundries and their customers are going to look for ways to ramp R&D and cost-sharing structures. There’s already been a great deal of speculation on whether or not a 28nm gate-last processor would have superior performance/power characteristics compared to Samsung’s 32nm design, but the answer to that question simply isn’t clear. It’s more accurate to say that either design could come out superior to the other, in the end, depending on a wide range of characteristics. Process node shrinks, in and of themselves, are no longer a guaranteed route to lower power consumption.
Apple A7 SoC
As for the A7 and 20nm ramping in the second half of 2013, it depends on your definition of the word “ramp.” In the x86 world, “production ramp” tends to signal a near-term launch. TSMC is a bit more fluid with such statements. In the past, the company has announced production on a new node a full year or more before products actually shipped. Even once a node has shipped, capacity constraints tend to keep practical product availability low for months thereafter.
With that said, TSMC is pushing hard. It’s even planning to try and jump forward to gain ground on Intel. All the same, it’s doubtful we’ll see 20nm shipping in volume before the middle of 2014, if not the beginning of 2015. With customers reportedly dragging their feet and loathe to commit to a node where the process and performance gains are expected to be smaller than any previous generation, there’s just not much reason to push the envelope on this one.