Simplicity makes for superfast computing 

 A radically new approach to computer design promises to deliver a supercomputer 500 times faster than any available today. Such a high-performance machine would be capable of performing more than 1 quadrillion operations per second.

 This week, IBM Research in Yorktown Heights, N.Y., announced a $100 million, 5-year exploratory research initiative to build such a computer. Nicknamed Blue Gene, it would be used initially to model how proteins fold themselves into the correct shapes to perform specific biological functions.

"The IBM announcement of its new research project is very exciting and important to high-end computing," says Thomas Sterling of the California Institute of Technology's Center for Advanced Computing Research in Pasadena. The project highlights innovative computer architecture as being crucial for rapid advances in computational power, he adds.

 The proposed machine would consist of about 1 million processors, which would share the computational load.

 Simplicity is key to the supercomputer's anticipated speed. "We use an ultraminimalist architecture for the processor design," says IBM's Monty M. Denneau.

 Processors in today's computers typically carry several hundred built-in commands. Most of those instructions, however, aren't actually used in many types of scientific computations.

 The IBM design cuts the number of instructions per processor down to a considerably more manageable 57. Moreover, each processor would be able to handle eight tasks at once instead of having to complete one task before going on to the next.

"Our goal was to reduce the size of the individual processor to almost nothing but to have a large number of them," Denneau says.

 Each of the computer's 32,000 microchips would hold 32 processors and 32 high-performance memory units for storing information and sharing it among processors. Keeping memory and processor close together should speed data access and greatly reduce power requirements.

 Even so, the computer, which would cover an area roughly the size of a tennis court, would consume about 1 megawatt of power and require a sophisticated cooling system.

 Denneau and his coworkers have also developed an innovative scheme for monitoring computations, checking for processor failures, and if necessary, redistributing the workload among still functioning processors on a chip. However, "it's going to take a couple of years to work out all the details," Denneau notes.

"It's a very interesting, revolutionary architecture," comments David V. Chudnovsky of the Institute for Mathematics and Advanced Supercomputing at the Polytechnic University in Brooklyn, N.Y.

 Even though fewer instructions are available, the simplicity should make the computer easier to program for applications ranging from modeling protein folding to performing fluid-dynamics calculations, Denneau says.