Replacing Wire With Laser, Sun Tries to Speed Up Data
Sun Microsystems is gambling on using beams of light to connect processor chips, eliminating a bottleneck.
Sun Microsystems is trying to do for computing what all the king’s horses and men failed to do for Humpty Dumpty. For decades, the semiconductor industry has broken silicon wafers into smaller chips to improve manufacturing yields.
Now Sun has found a way to reconnect the chips so they can communicate with each other at such high speeds that computer designers can build a new generation of computers that are faster, more energy-efficient and more compact.
The computer maker, which is based in Santa Clara, Calif., plans to announce on Monday that it has received a $44 million contract from the Pentagon to explore the high-risk idea of replacing the wires between computer chips with laser beams.
The technology, part of a field of computer science known as silicon photonics, would eradicate the most daunting bottleneck facing today’s supercomputer designers: moving information rapidly to solve problems that require hundreds or thousands of processors.
Processor and memory chips are currently made by etching hundreds or thousands of identical circuits onto a single wafer of silicon and then slicing that wafer into fingernail-size chips. That manufacturing process ensures that if there is a defect at a single spot on the giant wafer it will not ruin the entire batch of chips.
The drawback in the approach is that wires have to connect the chips in a computer. This causes a fundamental limit in processing power because data moves between chips at lower speeds, creating significant bottlenecks.
The wires that connect chips are analogous to the on and off ramps that cars use to move between freeways — just as cars slow down as they move onto city streets from multilane highways, electrical signals run more slowly between chips. The bottlenecks also generate additional electrical current and heat.
“All of a sudden it’s better to have an optical superhighway,” said Greg Papadopoulos, chief technology officer and executive vice president of research and development for Sun.
Computer scientists have long sought a way to make faster and cheaper computers by making larger chips on a single wafer of silicon, a manufacturing process called “wafer scale integration.” If the Sun researchers’ idea can be proved technically feasible and manufactured commercially, it would be possible to create more-compact machines that are a thousand times faster than today’s computers, the company said. Each chip would be able to communicate directly with every other chip in the array via a beam of laser light that could carry tens billions of bits of data a second.
The Sun researchers acknowledge that their project is a significant gamble.
“This is a high-risk program,” said Ron Ho, a researcher at Sun Laboratories who is one of the leaders of the effort. “We expect a 50 percent chance of failure, but if we win we can have as much as a thousand times increase in performance.”
Silicon photonics has become hot recently with major semiconductor and computer companies as well as start-ups investing heavily in efforts to build optical networking directly into processors to replace electrical wires.
Last week, the NEC Corporation, the Japanese supercomputer maker, announced that it had made an advance in optical connections between chips that will pave the way for a supercomputer able to reach speeds up to 10 petaflops, or 10 million trillion instructions a second. That is about 20 times faster than the world’s fastest computer.
Sun’s partners on the project are Stanford and the University of California, San Diego, and two silicon photonics firms, Luxtera and Kotura. The Sun bid was chosen over three competing teams from Intel and Hewlett-Packard; I.B.M.; and the Massachusetts Institute of Technology.
The five-year program is being financed by the Defense Advanced Research Projects Agency, and it builds on a Sun research project that was intended to interconnect chips electrically by stacking them edge to edge.
In 2003, a group of Sun researchers led by Ivan Sutherland, a computer industry pioneer, reported that they were able to transmit data inside a computer much more quickly than current techniques allowed. In an interview at Sun’s research laboratory last month, Jim Mitchell, a Sun vice president, said that the original technology was now moving closer to commercialization.
However, the new approach, which is based on the company’s ability to accurately align chips with high precision making it possible to transmit beams of light across the surface of the chips in ultranarrow channels called wave guides, could have a much bigger impact.
The Sun researchers refer to their new system as a “macrochip.” They said that the technology would make it possible for computer architects to completely rethink the organization of circuitry on a computer.
“It’s like the difference between having someone next door and having to get on an airplane to fly across the country,” said Alan Huang, an optical networking designer at the Terabit Corporation in Menlo Park, Calif. “This would be a way of breaking Moore’s Law.”
Moore’s Law is an observation by Intel founder Gordon Moore that the number of transistors on a computer chip doubles roughly every two years.
The Sun researchers said they believe that their alignment technique, which they plan to describe in a scientific paper, will make their systems more compact and easier to manufacture than their competitors’.