By forcing light to circle multiple times through ring-shaped structures carved into silicon, researchers at IBM have been able to delay the flow of light on a microchip. Being able to delay light is crucial for high-performance, ultrafast optical computers of the future that will process information using light and electrical signals.
It’s easy to store electronic data in computer memory; light is harder to control. The new silicon device, described in this week’s issue of Nature Photonics, is ten times smaller than those made in the past. It also works much better at high data speeds. “This work is approximately a factor of ten over the best achieved with [ring-shaped devices] so far,” says Keren Bergman, an electrical-engineering professor at Columbia University.
Storing light on silicon is key for electronic-optical hybrid computers that researchers believe will be available a decade from now. In these computers, devices will compute using electrons but will move data to other devices and components using light–avoiding the use of copper wires or interconnects that tend to heat up at high computer frequencies.
But the optical interconnects would have to be laid out in an intelligent network, just as the copper wires on today’s chips are. To transfer data packets efficiently between devices, the copper network on a chip has nodes where many interconnects converge. If a processor is sending data to a logic circuit, the data travels from node to node until it gets to the logic circuit. Each node in the network reads and processes the data packet to route it correctly to the next node. While the node makes a routing decision, it temporarily stores the data in electronic memory. To process and route data at the nodes of an optical-interconnect network, one would need to store, or delay, light so the node can make the routing decision.
Yurii Vlasov and his colleagues at IBM’s T.J. Watson Research Center delay light on a silicon microchip by circulating it 60 to 70 times through ring-shaped structures, called resonators. The researchers make these resonators on a thin silicon layer mounted on an insulting silicon-oxide layer. They etch parallel trenches into the silicon that reach down to the oxide. The raised portion between the trenches acts like a silicon wire that shuttles light.
The researchers employ the same silicon wafers and techniques that are used to fabricate microprocessors at IBM. This makes it easy to “think of combining optical circuitry with electrical circuitry on the same chip,” Vlasov says.
By connecting many rings, the researchers can build up the delay. With 56 rings connected to a common silicon wire, they get the longest delay: about half a nanosecond–which amounts to storing 10 optical bits–at a data speed of 20 gigabits per second.
Other researchers have made resonators on silicon before. But the smallest resonators so far have been about 100 micrometers wide, and cascading tens of them yields a device that is a few millimeters long–too big to be integrated into an electronic circuit. The IBM researchers make rings that are 12 micrometers in diameter, and they can fit up to 100 ring resonators into an area that is less than one-tenth of a square millimeter.
The size of the device is a major advance, Bergman says: “It is very close to the kinds of densities you would like to have on chip for optical interconnects.” Achieving a delay of 10 bits at gigabits-per-second speeds, which would be typical of the data speed that optical interconnects of the future would be handling, is a breakthrough, she says. “This is a major step towards making optical interconnects a reality.”
The device loses more light than would be acceptable in practical circuits, and Vlasov says that he and his team are working to reduce these losses. Once they do that, he says, they could put thousands of resonators together to store even more optical bits. For practical optical interconnects, you would need to store hundreds and thousands of bits.
It might take another 10 years before we see optical interconnects in computers, but the IBM research shows that the technology is viable, says Risto Puhakka, president of market-research firm VLSI Research, in Santa Clara, CA. “There are legs on this technology, and it could eventually be integrated with current circuits into chips.”