Will Tomorrow’s CPUs Give Off Light Instead of Heat?

There’s been a lot of talk about the convergence of computing and communications in recent years — a lot more talk than actual convergence. One reason or stumbling block has been that modern communications systems are increasingly optical, with beams of light flowing through fiber-optic tubes, while computing platforms are based on electrons flowing through copper or silicon wires.

Before long, however, new developments may finally combine the two, providing an optical bus for silicon CPUs, or perhaps even using light itself to conduct computations. This is the idea of optical computing or silicon photonics, and it’s moving ahead in both academic labs and chipmakers’ fabs.

“Microelectronics is now facing the problem that the overall delay in the processor is not related to the gate speed but to the wiring,” explains Lorenzo Pavesi, a professor of experimental physics at the University of Trento in the Italian Alps. This is the so-called “interconnect bottleneck.” In addition, PC and processor designers are already struggling with the power dissipated by silicon chips in the form of heat — power increasing at such a rate that, in a few years, CPUs could burn themselves up. Pavesi says, “Photonics will surely play a role in solving this bottleneck.”

Light is inherently more efficient than electricity. It is faster and can be multiplexed (a single fiber can carry multiple channels at different frequencies). And as anyone who’s put a Pentium 4 notebook on her lap can tell you, electrons moving through metal wires also generate a lot more heat than light moving through fiber optics.

The best kind of light for moving data is a laser. If your PC has a CD or DVD drive, there’s a tiny laser in your system already. To fulfill the promise of photonics, however, chipmakers will need to put a laser not just inside your computer, but inside a silicon chip.

Building a Silicon Laser

Lasers are relatively easy to make and quite common in the networking and telecommunications industry these days, but they’re far too big and expensive to fit on chips. To make photonics cost-effective on a chip level, manufacturers need to get the silicon itself to lase, or emit a coherent, focused beam of photons that can be switched on and off to transmit digital information. Unfortunately, silicon isn’t very well-suited to being a laser: It doesn’t emit light nearly as readily as currently popular materials such as gallium arsenide (GaAs).

“Silicon is an indirect-band-gap semiconductor — that is, light is not generated efficiently [compared to direct-band-gap GaAs],” Pavesi explains. And combining silicon and GaAs or another natural emitter on a single chip is difficult and expensive.

Several different strategies have been applied in the lab to make silicon emit light, including the use of silicon nanocrystals to yield LED-style visible light (pioneered by Pavesi’s team in 2000) and erbium-doped silicon –semiconductors doped with rare-earth ions — for 1.5-nanometer emissions. Pavesi says his and other researchers’ work proves that both these approaches are able not only to generate but to amplify light, or show optical gain, from silicon.

“What still has to be done is to implement these systems within an optical cavity [a ricochet array of mirrors that make the light coherent] and inject current though [stimulating the silicon with electricity rather than light] to have an injection laser,” Pavesi says. “I am very convinced that silicon nanocrystals at the end will be able to show lasing; other people are more convinced of the Er [erbium] approach.” STMicrolectronics, for example, claims it will have a Er-based silicon laser operational within the year.

Intel Sees the Light

Integrating photonics onto silicon chips isn’t just an academic exercise. Intel Labs is conducting massive amounts of research into silicon photonics, spending some $4 billion in R&D; in 2002 alone. Intel vice president and chief technology officer Pat Gelsinger first showed the work’s importance to future chip designs at the Spring 2002 Intel Developer Forum Conference, where he demonstrated a tunable optical filter built directly in silicon.

Tunable laser filters can quickly modulate optical wavelengths to increase data rates. Although only a prototype, Intel’s optical chip used only $1 worth of components. And because it was programmable, it could be adjusted to filter different frequencies via software.

Paging Peter Parker

Although the first opto-computers will probably use fiber-optic cabling, researchers are working on creating more advanced optical channels. One group at the University of California at Riverside has found a way to make superfine photon passageways using threads of spider’s silk.

The fibers are made much like old-fashioned candles: The scientists dipped thread from a Nephila madagascariensis, which is indigenous to Madagascar, into a solution of tetraethyl orthosilicate and allowed the solution to dry. They then heated the fibers, which burned away the silk and caused the surrounding tube to shrink to a diameter of just one micrometer. By using thinner silk, the team hopes to create tubes just two nanometers in diameter, which could be very effective for carrying photon flows.

The device meaures just a few microns wide by a couple of millimeters long, and is able to filter wavelengths in the DWDM (Dense Wavelength Division Multiplexing) spectrum — where today’s optical filters, which can cost upwards of $10,000 each, manage and select among the different frequencies of laser light in a communications channel. “There is a different optical defraction characteristic for light based on either the bias or temperature of silicon,” Gelsinger explained. “You actually change this defraction index to select different band-pass frequencies.”

At the same conference, Gelsinger predicted that eventually Intel will be building all-optical silicon chips that process light waves instead of electrons. “We want to take optics from the WAN, to the enterprise, to the LAN, to the data centers, to the rack, and finally from chip to chip,” he said. “Today, making an optical connection costs many tens of thousands of dollars. Our job is to make that pennies, resulting eventually in the opto-processor — a microprocessor with direct optical interfaces.”

An Evolutionary Revolution

Despite the potential, even believers like Pavesi see huge challenges in building an all-optical processor for computation. Pavesi thinks optical transistors will be difficult to deploy in large numbers, but that closing the gap between optics and electronics is still possible: “What I belive is that photons will be used in processors not to make computations, but to transfer information.”

Agilent Labs, formerly part of HP’s R&D; division, is working on a range of opto-electric devices, including some designed specifically for computation. In late 2002, Waguih Ishak, director of Agilent’s Communications & Optics Research Laboratory, said the company was developing an “on-off crystal switch” that could work as an optical transistor. A collection of these programmable photonic crystals could evolve into an all-optical CPU. Agilent is also working on creating an all-optical bus system for moving data around a system.

When this optical bus is finished, the chip could have a direct connection to a fiber-optic cable, essentially bridging the opto-electric gap and drastically increasing communications bandwidth. Once this optical bridge is laid, data processors could truly hit light speed.

Categories: Technology