The Route To Finer Lines

Chilly Crystals

Every few years scientists and manufacturers from all over the globe draw up what is now called the International Technology Roadmap for Semiconductors, an assessment of semiconductor technology requirements and research goals over the next 15 years. Ironically, one of the biggest challenges the industry faces is traffic congestion on and between the chips themselves.

Thanks to ever shrinking transistors on integrated circuits (ICs), computers have become quicker and more powerful. But as faster and smaller transistors are packed onto a microchip, the layers of wires that connect the transistors must shrink as well. The problem, though, is that the smaller the cross section of a wire, the tougher it is to push an electrical signal through. Capacitance between extremely thin wires can add to the trouble. "The transistors are getting faster, but the wires are getting slower--and that's a prescription for disaster," says Kevin Martin of the Georgia Institute of Technology, who helps to direct the Interconnect Focus Center, an entity created to avert that disaster. Based at Georgia Tech, the center encompasses research at five other universities and is part of the larger Technology Focus Center research program, launched in 1998 with $10 million annual funding per center from the Semiconductor Industry Association's member companies and other groups.

The semiconductor industry has high hopes for the interconnect program. Right now the best commercial interconnects are copper wires, introduced into microchips in late 1998. But although copper is a vast improvement over aluminum interconnects, Martin says, the metal simply won't scale down sufficiently. For example, the intrinsic switching time for transistors having 100-nanometer gate lengths (circuit features referring to distances that electrons must travel) is on the order of 0.1 picosecond, 70 times faster than the response time of a typical one-millimeter-long copper interconnect wire. And the pressure is on--the International Technology Roadmap calls for chips with 100-nanometer gate lengths next year.

Leading the race for new interconnects are optical ones--replacing wires with fiber-optic cables that are resistance-free. Optics are ideal for high-bandwidth applications and are not constrained by long distances, unlike wire interconnects. Research at the Massachusetts Institute of Technology is focusing on sending signals between transistors on the chip itself, whereas David A. B. Miller, an electrical engineer at Stanford University, has directed his efforts at enabling separate chips to talk to one another at the necessary speed without having to be crammed closely together. "Using optics instead of wires is like being able to put in a 1,000-lane highway where you previously had a one-lane freeway," Miller remarks.

There are two main approaches to optical interconnects, albeit with myriad variations. One is transmitting light beams, generated by five-to 20-micron-high vertical cavity-surface-emitting lasers, or VCSELs, down waveguides built onto the chips. The other paradigm is based on freespace optics. Light from an external source can be reflected by tiny structures called quantum-well light modulators, which rapidly switch on and off in response to small voltages. Alternatively, patterns of light generated on one chip by VCSELs can be imaged on the other chip by a lens. "The second chip behaves like your retina," Miller explains. Though not yet ready for prime time, optical interconnects have been successfully demonstrated at several universities.

Just out of the gate, so to speak, is wireless-interconnect technology using radio-frequency (RF) signals. Various groups are working on this concept, including M. C. Frank Chang of the University of California at Los Angeles under the auspices of the Interconnect Focus Center. One example of how RF interconnects would work was presented in March by Kenneth K. O of the University of Florida and graduate students Brian A. Floyd and Kihong Kim at the International Solid-State Circuits Conference. They delivered a paper on the use of RF signals in massively parallel computers. With such large computer systems, maintaining a constant clock signal throughout numerous microprocessors becomes difficult. O's group hopes to get around that by broadcasting a clock signal from one IC to others using microwaves. One design integrates millimeter-size receivers and antennae on each IC in a multichip module. "By propagating the signal at the speed of light, we're trying to reduce the clock skew," O says. "You could send a wave down to a multichip module and provide equal clock phase to a very large area."

The group recently demonstrated on-chip wireless transmission and reception of a 7.4-gigahertz clock signal. O believes the same technology could be modified for data transfer between chips as well. Not surprisingly, the biggest antagonist to wireless interconnects is noise. Both the chip's silicon substrate and the switching of the transistors themselves degrade and taint the radio signal. The materials in chips "are just not very friendly to radio reception," O says. Whether optics or RF, researchers will undoubtedly find ways to keep traffic moving on tomorrow's computer systems.

----David Pescovitz