The Turn: Fabrication
The reason ARM is dominant on the mobile side is that, to date, Intel has been unable to demonstrate a power-efficient MSoC. In this world, trumpeting impressive performance-per-watt numbers isn’t enough. You actually have to be able to show off a full day’s worth of talk time and impressive standby numbers in order to be functional. With Medfield, Intel demonstrates that its team has the technical know-how to produce an MSoC within striking distance of ARM. As Intel put it, Medfield buys the company a seat at the table.
You hear a lot about the relative pros and cons of the ARM and x86 architectures, and it is true that Qualcomm and ARM have more success with low power consumption. Until Medfield, no one was sure how much better-off those companies were compared to Intel. Now we can say that they had a four-year head start, since it took Intel that long after the launch of Atom to come up with a competitive mobile platform.
What happens in the next three years, though? Given that Medfield is competitive with currently-shipping ARM MSoCs, we have to look to the next generation. On the previous page, we suggested that ARM and Qualcomm face at least as significant of a challenge scaling performance up as Intel faces in scaling power down.
We can be perhaps most objective in looking at manufacturing technology. Intel has the best chip fabs in the industry, which allowed it to out-compete AMD during the K6 and K7 era, and maintain its position when AMD introduced the successful K8-era processors. Medfield is currently based on a 32 nm node and is already competitive with ARM-based solutions. Intel’s next move is to make a jump to 22 nm on a 3D FinFET design, representing two steps forward in process technology. Intel has never failed to execute with a fabrication process, and it will already have plenty of experience from its Ivy Bridge-based processors. If the company stays on track, it’s about 18 months ahead of the competition in manufacturing. As soon as the competition starts shipping 28 nm, Intel will follow with 22 nm, and it will be even longer before competing fabs can implement FinFET. This gives Intel another 20-30% improvement in power consumption over its current technology, while basically doubling density.
High-K/Metal Gate
ARM-based vendors are also in a race to enable MSoCS manufactured with high-K/metal gate technology. Besides Intel and Samsung, all of the other players are dependent on outside foundries like IBM, Globalfoundries, and TSMC. Samsung opens its foundries to other companies, so they’re on that list too.
Qualcomm signed on with Globalfoundries to manufacture its 28 nm MSoCs. Globalfoundries uses a “gate-first” high-k/metal gate 28 nm process. This is also what IBM and Samsung will be doing. They won’t attempt to make the switch to “gate-last” until 20 nm lithography is available. TSMC, the original foundry-for-hire (and the expected manufacturing partner for Apple’s A6) is going with a “gate-last” approach. Intel has always gone with a gate-last approach.
“Gate-first” and “gate-last” describe the way you implement modern high-k insulators with metal gates, and the layman’s explanation is that we’re talking about a shift from pure silicon dioxide-based chips to those containing a different insulator called hafnium. In a gate-first design, you put the metal gate and hafnium components on the wafer before heating it, and in a gate-last design (like Intel’s) you do the replacement afterwards.
Gate-first buys you more density, which is great for performance, but it comes at the expense of yield. You have more failures in manufacturing. Gate-last is more reliable for manufacturing. But there are more restrictive design rules that you have to get around. Intel has been shipping high-k chips since its Penryn core in 2007. Globalfoundries didn’t successfully ship high-k chips until 2011.
Intel isn’t wrong for choosing gate-last. It has a proven track record with its Penryn, Nehalem, and Sandy Bridge architectures. Qualcomm may be wrong for choosing gate-first. Globalfoundries has a working gate-first process, used to manufacture AMD’s APUs. Unfortunately, as noted in AMD’s third quarter earnings report, yields weren’t as good as previously hoped, leading to lower-than-expected revenue growth.