Importantly, TSMC’s decision to go with gate-last is steeped in history, according to the company’s senior VP in charge of R&D. Part of the reason why gate-first manufacturing results in low yields is that you have to control threshold voltage carefully, since the N- and P-channels use the exact same metal. The semiconductor industry tried to carefully control the voltage this way two decades ago and found it very difficult. The gate-last approach doesn’t require the same control because the metal for the P channel is different than the metal for the N channel. You lose some density, but yields are a lot higher, and the easiest way to lose a fight is to not show up at all. It’s not trivial to switch a design from gate-first to gate-last. It requires additional redesign time. To that end, you can’t just change your order from Globalfoundries to TSMC by checking a different box on a form.
It seems that Qualcomm figured out it can’t get the yields it needs on a gate-first approach. At the 2010 International Electron Devices Meeting held in San Francisco, the company stated that it wouldn’t be using high-k/metal gate technology for the majority of its 28 nm products. This is a big disadvantage for Qualcomm.
Intel’s 32 nm high-k design (Medfield) is competing favorably against current 40/45 nm ARM-based CPUs.
In the next iteration of its product, Intel will jump from 32 nm to 22 nm FinFET, equivalent to two process jumps. Qualcomm is going from 45 nm to 28 nm (1.5 nodes). But it isn’t able to make a jump to high-K, translating to a loss of clock rate and a power consumption sacrifice. Apple and Nvidia are both expected to rely on TSMC for their next-generation chips. Apple will go from 45 nm to 28 nm high-k (1.5 nodes) and Nvidia will go from 40 nm to 28 nm high-k (1 node). With 28 nm high-k, both Apple and Nvidia should have extra performance compared to Qualcomm’s jump to 28 nm silicon.
All things equal, Intel gets the biggest boosts to power and performance from process technology advances. Qualcomm, gambling on gate-first 28 nm, would have enjoyed more chip density if yields were good enough. Now, though, it’s forced to back off and go with a silicon-dioxide 28 nm process.
But all things aren’t equal. In order to compete, Qualcomm has to roll out new tricks like out-of-order execution and an improved memory architecture. The company’s engineers are doing this for the first time with Krait. They have to be flawless in their execution. Intel, on the other hand, is already doing well with a relatively low-tech Atom that doesn’t incorporate any of the advancements seen from many, many years of designing x86 processors.
Of course, Intel’s process technology lead will continue moving forward. While everyone else tries catching up to the firm’s high-k/metal gate manufacturing, it has already publicly demoed Claremont, a Near-Threshold Voltage Processor that operates at less than 10 mW. Interestingly, this chip was built around the original Pentium core, much like the Atom.
So, Intel leads with regard to architectural challenges and on the manufacturing side. There’s just one more piece of the puzzle: graphics.