As the trend famously codified by Intel co-founder Gordon Moore -- that the number of transistors on an integrated circuit would double every two years -- seems to be flagging, one top engineer suggests that it is time to rethink chip design to buy performance increases.
Instead of just focusing on reducing chip size and cranking up processor speeds, engineers need to look at making tweaks or possibly even change fundamental microprocessor architecture to ensure chips are faster and cheaper to produce, said Robert Colwell, director of the microsystems technology office at DARPA (Defense Advanced Research Projects Agency), in a talk on Monday at the Hot Chips conference in Stanford, California.
[ Keep up on the day's tech news headlines with InfoWorld's Today's Headlines: Wrap Up newsletter. ]
Colwell dismissed arguments that Moore's Law will continue to hold, and said engineers should give serious thought to the design and economics of chip making.
"I pick 2020 as the earliest ... when we can call it dead. That's only seven years away," Colwell said. "I'm thinking 7-nanometers. You can talk me into 2022, you might be able to even talk me into 1-nanometer. But you're not going to into 1-nanometer ... I think physics dictates against that."
There have been different interpretations of Moore's Law, with the most common one being that the number of transistors on a chip will double every two years, which will make chips faster. But Colwell tried to clarify the definition, saying that in 1965 Intel's Gordon Moore focused more on the economics related to cost-per-transistor, which would drop with scaling.
"What it really is ... if you're going to integrate a lot of components on a chip, there's an optimal place on that curve where you should do that. You can go beyond the optimal point, but it will cost you in terms of price per component," Colwell said, adding there's a sweet spot where maximum profit can be eked out assuming sales are relative to the number of chips made.
It is true that beyond a certain geometry it will hard to make chips smaller, but Colwell said economics, and not physics, would ultimately end Moore's Law. The day chip makers can't get return on the billions invested in making chips smaller is the day Moore's Law will break. Instead of waiting for chip economics to crash, innovation should start now.
Chips now have billions of transistors, and the ability to push clock speed and performance will reached its limit, and after silicon engine stalls, small tricks and incremental tweaks like power gating and turbo will happen for a short time to improve chips. But chip designers should start early, and an effort like changing fundamental chip design could help before and after Moore's Law ends.
One approach to consider might be separating the instruction set architecture, microarchitecture, circuits, functional blocks and other parts now integrated onto a chip, and tweaking them for specific applications, Colwell said.
"I think the end of Moore's Law opens the door to designing special-purpose things again," Colwell said, adding that in the 1970s, one could make specialized floating point arrays with vector processors. DARPA is doing research in the areas of quantum computing, nanotechnology and distributed computing, he said.
Researchers at universities and chip companies are also looking at new materials to replace silicon, and also advanced manufacturing technologies. Colwell said that the new technologies are far from practical implementation, and chip makers will have to rely on technologies like CMOS, which has no practical replacement in sight.