Chip Makers Fight Dwindling Gains in Efficiency
The latest generation of graphics chips have 3 billion transistors and consume about 200 watts of energy. The numbers are impressive -- until you consider that the human brain has the equivalent of a trillion transistors and consumes just 20 watts of energy, or far less than it takes to run a light bulb.
Semiconductor makers are looking at the brain with envy these days as they deal with the latest challenge to their industry -- dwindling gains in power efficiency. It's long been a concern for chip designers, but it's taking on new urgency as the common techniques of scaling down power usage are losing their effectiveness.
"The set of factors that have worked for us for about the last decade seems to be scattering around the edges," said Jan Rabaey, a professor at the University of California, Berkeley's College of Engineering, who moderated a panel on the topic at the Solid-State Circuits Conference in San Francisco this week.
It's not an abstract concern, either. Steady improvements in power efficiency have been a key enabler of today's powerful computers, especially mobile devices like the iPhone, where battery life is crucial.
The biggest gains have come from "process shrinks," or the move to new manufacturing techniques that allow for smaller and smaller transistors. It's the regular advance that's best known for enabling Moore's Law, but it has also allowed the performance per watt of semiconductors to improve with each process generation.
Process shrinks historically gave a 3x boost in energy efficiency, but today's advances give only a 1.4x improvement, said Dan Dobberpuhl, a former chip engineer at Digital Equipment Corp., Broadcom and Apple.
"Below 30 nanometers we have to introduce new materials and new structures" to keep scaling transistor voltage down, Oh-Hyun Kwon, president Samsung Electronics, said earlier in the day.
Engineers have been employing other tricks to reduce power consumption, such as controlling leakage, but the returns there are getting smaller too. The panelists were asked for proposals to bring about "the next 10x reduction" in power usage.
New transistor designs are part of the answer, said Jack Sun, CTO at contract manufacturing giant TSMC. Options include a design called FinFET, which uses multiple gates on each transistor, and another design called the junctionless transistor.
Researchers have made "great progress" with FinFET, and TSMC hopes it can be used for the next generation of CMOS -- the industry's standard silicon manufacturing process, Sun said.
He and other panelists also have faith in a packaging technique dubbed 3D stacking, in which chips are layered on top of each other instead of side by side. It allows for shorter interconnects to bind them together, reducing electrical waste.
Many of these techniques are still at the research stage, however. And if they are not compatible with current CMOS manufacturing equipment they will be expensive to implement.
There's another way to come at the problem. Chips are fairly inflexible today, in the sense that they don't adapt much to their environment, said Philippe Magarshack, vice president for research and development at STMicroelectronics. He proposed an approach he called "sense and react."
The demands placed on a smartphone chip can vary widely. Chips should be able to ramp their voltage, clock speed and other properties up and down depending on whether a phone is making a call or showing a video, or whether it's close to a base station or far away, he said.
It's done to an extent today with clock gating and voltage scaling, but it could be done much more if each component -- the antenna, receiver and so on -- were designed in concert, Magarshack said.
"The holy grail is that the system only expends power and energy when needed. This is definitely not possible with the tools we have today. We need the next generation of tools and methods," he said.
He and Sun also proposed a so-called wide I/O architecture, which would allow multiple components to share an input/output device. A DRAM chip could be stacked on a baseband processor, for example, and do parallel instead of serial I/O operations, Magarshack said.
Dobberpuhl, the former DEC engineer, said the biggest gains can be made from improved algorithms and architectures, including more efficient parallel designs.
Herman Eul, president of Intel Mobile Communications, said the key is moving functions handled by analog processors into digital. Digital circuits are easier to shrink, he said, and they can be reprogrammed -- so a single transceiver could be used for all five frequency bands in a 3G phone, for example, instead of the five separate chips used today.
"The most power effective transistor is the transistor which is not there," Eul said.
In general the panelists seemed optimistic, if only because engineering persistence has broken through many barriers in the past.
"In general, engineers never give up," Eul said.