SANTA CLARA, Calif. – FinFETs hold both promise and peril and are not yet ready for prime time, according to a panel of experts who spoke on their experiences with the technology at the annual Synopsys User Group meeting here.
The 3-D transistor structures coming at the 14-nm node promise either performance increases or power consumption reductions of more than 60 percent compared to today’s 28 nm process, according to a Globalfoundries technical executive. However, they increase gate capacitance on a per micron basis, raising a handful of old and new design issues, said others on the panel.
FinFETs bring a 66 percent increase in gate capacitance per micron compared to today’s-28 nm process, back up to the level of the 130-nm planar node, said Anil Jain, vice president of IC engineering at Cavium Networks. Capacitance will limit both performance increases and dynamic power scaling for high-end chips, he said.
“We have these beautiful [3-D] transistors, but we can’t run them too far,” Jain said, noting “dynamic power [is] getting out of hand.” In addition, “those of us in high performance devices have not seen much improvement in core voltage scaling,” he said.
Click on image to enlarge.
Cavium estimates a 40 percent increase in gate capacitance due to FinFETs.
Jain asked EDA vendors for tools that do a better job controlling switching power and isolating electro-migration faults. “FinFETs are not a no-brainer migration--at the end of the day we have to pay for it, so please don’t blow up our costs,” he added.
Michael Campbell, a vice president of engineering in Qualcomm’s chip design group, said FinFET structures in one foundry are “similar, but not the same” as those in another foundry. “You can only etch in certain directions and etch tools are common—that drives some similarities--but [foundries] use different in tricks in spatial walls and diffusion,” he said.
Campbell noted pictures of Intel’s 22 nm FinFETs show irregularly tapered walls that could impact planar fault models. “It takes new test technologies and incredibly deep partnerships to do design for test properly,” he said.
The "gate capacitance" is no longer just the inversion gate capacitance. In fact, inversion gate capacitance is just a small portion of the total transistor capacitance. FinFET has significantly higher parasitic capacitance than planar devices, and that's just one of the few reasons even Intel FinFET did not the 50% reduction of active power that was originally claimed.
Anil Jain said “ FinFETs bring a 66 percent increase in gate capacitance per micron compared to today’s-28nm process, back up to the level of the 130-nm planar node”. I disagree. At the device level the transistor drive current is improved by high gate capacitance resulted from higher permittivity of the high-k dielectric. However, performance at high operating clock frequency can be negatively impacted by the increased gate capacitance. To minimize this penalty Intel uses the unique HK/MG (high K such as HfO2)/Metal electrode that give intrinsically superior electrostatic control or short channel effect and reduced stand-by power reducing gate current. Therefore, Intel’s HK/MG will provide high performance through higher gate capacitance and concurrently scale down the gate length of the transistor. That is why Intel’s 22nm FinFETs are in high volume manufacturing over a year, and the 14nm FinFETs will be manufactured in 2014. Recently, major foundries, TSMC, Samsung, and others announced to adopt FinFET technology at 14nm and manufacture in 2014, skipping the 22nm to catch up with Intel. However, the successful implementation of FinFET technology will be much easier at the 22nm than at the 14nm node because of the lack of process and manufacturing learning at the 22nm node, resulting in further behind Intel.
Valid concerns with the increased capacitance, however that only points out a potential increase in dynamic power. Recently, leakage power has been of major concern in Deep Submicron Planar technologies. This power is consumed while the gates are not switching and historically have contributed much of the power. Perhaps finfets offer the benefit of reduced leakage power.....
I guess I should add intel BS marketing as well.
Paul O. CEO has been talking about a finfet and winning mobile for past 3 years. yet intel still not shipping mobile atoms using finfet. In fact Paul will be pushed out of intel before a singe cell phone or table chip ship with finfet ships.
I hear uncompetitive 22nm finfet atoms too high power and shipping slipping to 2014 . I think Rick you found the problem. Capacitance is too high and finfets increase not lower power .
40% or 66 % higher capacitance "equal to 130nm "and little supple voltage scaling !
Thus Power can not be lower based on physics Power being proportional to C and V*V . I am loosing all respect for TSMC and GF marketing.
Rick or anyone else any comment?
David Patterson, known for his pioneering research that led to RAID, clusters and more, is part of a team at UC Berkeley that recently made its RISC-V processor architecture an open source hardware offering. We talk with Patterson and one of his colleagues behind the effort about the opportunities they see, what new kinds of designs they hope to enable and what it means for today’s commercial processor giants such as Intel, ARM and Imagination Technologies.