Transistor announcements aren’t the sexiest of occasions, but Intel’s 22nm SoC unveil is important for a host of reasons. As process nodes shrink and more components move on-die, the characteristics of each new node have become particularly important. 22nm isn’t a new node for Intel; it debuted the technology last year with Ivy Bridge, but SoCs are more complex than CPU designs and create their own set of challenges.
Like its 22nm Ivy Bridge CPUs, the upcoming 22nm SoCs rely on Intel’s Tri-Gate implementation of FinFET technology. According to Intel engineer Mark Bohr, the 3D transistor structure is the principle reason why the company’s 22nm technology is as strong as it is. Other evidence backs up this point.
Earlier this year, it emerged that Nvidia was deeply concerned about manufacturing economics and the relative strength of TSMC’s sub-28nm planar roadmap. Morris Chang, TSMC’s CEO, has since admitted that such concerns are valid, given that performance and power are only expected to increase by 20 to 25 per cent as compared to 28nm.
Intel, in contrast, is predicting record gains. The company claims that its 22nm SoC “employs high speed logic transistors, low standby power transistors, and high-voltage tolerant transistors in a single SoC chip to support a wide range of products, including premium smartphones, tablets, netbooks, embedded systems, wireless communications, and ASIC products.”
The company reports enormous improvements in leakage currents and Intel plans to take full advantage of the improved performance.
You’ve probably seen the image above trotted out when Intel talks about process node improvements. In this case, it’s the length of the line that’s more of an improvement than its rightward shifts. The diagram shows leakage current dropping more quickly than clock speed. At 65nm, Intel’s transistor performance and minimum leakage levels dropped off more quickly, while minimum leakage was much higher.
Here’s 65nm, 32nm, and individual data sets for SRAM cells across multiple process nodes.
At 65nm and a maximum input voltage of 1V, Intel’s SRAMs had a narrow operating range. 800MHz was the maximum effective frequency at that voltage – below 0.8V, the chip stopped working at any frequency. At 32nm (Medfield, Clover Trail), the company’s processors have considerably more latitude. 22nm pushes the envelope still further.
The challenge for both TSMC and GlobalFoundries is going to be how to match the performance of Intel’s 22nm technology with their own 28nm products. 20nm looks like it won’t be able to do so, which is why both companies are emphasising their plans to move to 16nm/14nm ahead of schedule. There’s some variation on which node comes next; both GlobalFoundries and Intel are talking up 14nm, while TSMC is implying a quick jump to 16nm.
I don’t want to say too much on how the three companies’ future processes might compare – tech papers at IEDM may shed more light on the particulars of each solution. What’s clear is that both GF and TSMC are going to try to accelerate FinFET development. GF’s tech papers imply that the company will deploy a hybrid 22nm-14nm process to make the jump more quickly.
Will it work? That’s an unknown at this point. TSMC and GlobalFoundries both have excellent engineers, but FinFET is a difficult technology to deploy. Ramping it up more quickly than expected while simultaneously bringing up a new process may be more difficult than either company anticipates. Given the advantages Intel claims for the technology, it might have made more sense to ramp FinFET on an established node.
One of the most significant demonstrations of what Intel thinks it’s getting out of 22nm FinFET is the company’s decision to revise Atom for an out-of-order architecture. Intel has resisted the call to overhaul the in-order CPU; the current core at the heart of Medfield and Clover Trail offers nearly identical performance to the design that debuted in 2008.
22nm Atom should close the gap with existing ARM CPUs and give Intel a substantial advantage. Overall, the situation looks like Intel holds the cards until GF and TSMC manage to revise their roadmaps for the sub-20nm market.