Beyond silicon: $1.5 billion U.S. program aims to spur new types of computer chips
Silicon computer chips have been on a roll for half a century, getting ever more powerful. But the pace of innovation is slowing. Today the U.S. military’s Defense Advanced Research Projects Agency (DARPA) announced dozens of new grants totaling $75 million in a program that aims to reinvigorate the chip industry with basic research into new designs and materials, such as carbon nanotubes. Over the next few years, the DARPA program, which supports both academic and industry scientists, will grow to $300 million per year up to a total of $1.5 billion over 5 years.
“It’s a critical time to do this,” says Erica Fuchs, a computer science policy expert at Carnegie Mellon University in Pittsburgh, Pennsylvania.
In 1965, Intel co-founder Gordon Moore made the observation that would become his eponymous “law”: The number of transistors on chips was doubling every 2 years, a time frame later cut to every 18 months. But the gains from miniaturizing the chips are dwindling. Today, chip speeds are stuck in place, and each new generation of chips brings only a 30% improvement in energy efficiency, says Max Shulaker, an electrical engineer at the Massachusetts Institute of Technology in Cambridge. Fabricators are approaching physical limits of silicon, says Gregory Wright, a wireless communications expert at Nokia Bell Labs in Crawford Hill, New Jersey. Electrons are confined to patches of silicon just 100 atoms wide, he says, forcing complex designs that prevent electrons from leaking out and causing errors. “We’re running out of room,” he says.
Moreover, only a handful of companies can afford the multibillion-dollar fabrication plants that make the chips, stifling innovation in a field once dominated by small startups, says Valeria Bertacco, a computer scientist at the University of Michigan in Ann Arbor. And some big companies are going down separate paths, designing specialized chips for specific tasks, Fuchs says. That has reduced the incentive for them to pay for shared, precompetitive basic research. The number of companies involved in the Semiconductor Research Corporation in Durham, North Carolina, which backs such work, dropped from 80 in 1996 to less than half that in 2013, according to a study by Fuchs and her colleagues.
DARPA is now trying to fill the gap, with grants to researchers such as Shulaker. He is fashioning 3D chips with transistors made of carbon nanotubes, which switch on and off faster and more efficiently than silicon transistors. Companies today already make 3D chips with silicon as a way to pack logic and memory functions closer together to speed up processing. But the chips are slowed down by bulky and sparse wiring that carries information between the chip layers. And because 2D silicon chip layers must be fabricated separately at more than 1000°C, there is no way to build up 3D chips in an integrated fabrication plan without melting the lower layers.
Carbon nanotube transistors, which can be made nearly at room temperature, offer a better path to dense, integrated 3D chips, Shulaker says. Even though his team’s 3D chips will have features 10 times larger than state-of-the-art silicon devices, their speed and energy efficiency is expected to be 50 times better—a potential boon for power-hungry data centers.
The DARPA program is also supporting research into flexible chip architectures. Daniel Bliss, a wireless communications expert at Arizona State University in Tempe, and his colleagues want to improve wireless communications with chips that can be reconfigured on the fly to carry out specialized tasks. Bliss is working on radio chips that mix and filter signals using software rather than hardware—an advance that would allow a larger number of devices to transmit and receive signals without interference. This could improve mobile and satellite communications, as well as enable a rapid growth in the internet of things, where myriad devices communicate with one another, he says.
Another DARPA grant, for researchers at Stanford University in Palo Alto, California, will go to improving computer tools used in chipmaking. These tools verify novel chip designs with a form of artificial intelligence called machine learning. They would help automate the largely manual process of detecting design bugs in chips made up of billions of transistors, and could speed up the ability of companies to design, test, and fabricate new chip architectures.
If even a fraction of the new projects succeed, the DARPA project “will completely revolutionize how we design electronics,” says Subhasish Mitra, a Stanford electrical and computer engineer, and a researcher on the 3D carbon nanotube and circuit validation projects. He says it will also spur engineers to look beyond silicon, which has dominated research for decades. “When I was a student, life was boring,” Mitra says. “It was clear that silicon would move forward along a known path. Now, it’s absolutely clear that’s not what the future is.”