For the microprocessor industry, the future is one word: optics.
That's where the likes of Intel and IBM expect to find the revolution that will keep accelerating and miniaturizing chips in accordance with Moore's Law into the terahertz age and beyond -- long past what electronics alone can do. The age of gigabytes-per-second downloads and PDAs with the power of today's server farms will dawn only when photons can shoulder some of the work now handled by old-fashioned electronics.
And at this week's annual Photonics West conference in San Jose, California, two breakthroughs will be announced that advance the prospects for hybrid chips that compute using both electronics and optics.
Today, of course, computer chips are entirely electronic, while wide-scale computer networks (such as the internet) are mostly optical. Copper connects the components that make up a computer; copper connects local-area networks; fiber optics connect things beyond that.
Now, copper is increasingly proving incapable of shuttling bits across even short distances with the rapidity demanded by faster clock speeds.
"Similar to the telecom industry, (optics in computers) have been going from the longer distances into the shorter distances," said Marc Taubenblatt of IBM's Thomas J. Watson Research Center. IBM, he said, has been using optics to connect across "machine-room distances" -- between, say, a mainframe and a storage system -- since 1990 (.pdf).
The battleground right now, he said, is in connecting server racks together. "Optics is winning the rack-to-rack competition with electrical," said Taubenblatt. As soon as 2010, connections between cards, or "blades," on a rack will go optical, followed by components on a single motherboard.
That leaves communication within a chip as the final domino.
"If you look out in the middle of the next decade, when (processors will contain) hundreds of cores, you're looking at terabits of (on-chip) communication right there," said Mario Paniccia, director of Intel's Photonics Technology Group. "That's very difficult to do with copper."
That's when, maybe 15 or 20 years out, electrons will nearly exclusively be the stuff that computes, while photons will nearly exclusively be the stuff that communicates.
And ideally, it'll still all be done on good, old-fashioned silicon chips -- so computer manufacturers won't have to waste the billions of dollars invested in facilities building conventional computer chips, called complementary metal-oxide semiconductors, or CMOS.
That's where this week's announced breakthroughs come into play.
Optical communications on a computer chip require the mastery and miniaturization of three basic components: one that encodes an electrical stream of bits into light pulses (either using an on-chip laser or a modulator that acts like a shutter for laser light generated outside the chip), a conduit that pipes that light signal to its destination, and a receiver that decodes the optical bits back into an electrical signal.
Significant advances in parts one and three will be announced this week at Photonics West.
Paniccia's group at Intel will announce its fabrication of an optical modulator on a silicon chip that can translate electronic signals to light at speeds up to 20 GHz. That's nearly a threefold speedup from the group's previous modulator. The group's paper detailing this discovery is in this week's issue of the online journal Optics Express.
Expensive optical modulators have already been built of exotic materials, such as the crystalline molecule lithium niobate. But nothing is more friendly to mass production than silicon.
Andy Knights, of the Department of Engineering Physics at McMaster University in Hamilton, Ontario, notes that the Intel group's new silicon-based modulator "approaches that of the fastest commercial devices available, such as those fabricated using (lithium niobate)."
Slightly less difficult -- although still challenging -- is part three of the optics equation: making sub-millimeter-size detectors to convert the optical pulses back to electrical signals.
M.W. Geis and collaborators from MIT's Lincoln Laboratory will be announcing a breakthrough in that area this week as well: an all-silicon 10- to 20-GHz detector that, as it happens, can keep up with Intel's new modulator.
Their discovery will be published in the Feb. 1 issue of the journal IEEE Photonics Technology Letters.
"Integration of these devices with CMOS microelectronics is potentially straightforward," said McMaster University's Knights. "It's truly an exciting time in silicon photonics at the moment."
IBM has taken the lead in the middle component of the triad, microscopic silicon waveguides to pipe the information-carrying photons from the laser/modulator to the detector on the other side of the chip.
In December, Yuri Vlasov and colleagues from IBM published in the journal Nature their development of micron-size optical tracks that included storage rings. The latter devices would be used like miniature racetracks for the photons to circle around until the information they carry is needed.
These optical buffers managed to hold light for up to 60 laps around the track -- setting the buffered light pulses 10 bits behind unbuffered light.
"That's a record big number," said Vlasov, although the requirements of a typical microprocessor environment involve buffering of "hundreds of thousands of bits."
All the same, the prospect of integrated silicon chips containing both micro-optics and microelectronics has come much closer to reality in just the past few years.
Last September, Intel's Paniccia and John Bowers of the University of California at Santa Barbara announced that they'd invented a microchip-based laser composed of both silicon and the semiconductor indium phosphide. Until then, "We could do everything in silicon except for the laser," Paniccia said.
"We've proven that we can build devices in silicon that can be optical-friendly," said Paniccia. "Three years ago, everyone thought we were nuts."
