Computing at the Speed of Light; November 2004; Scientific American Magazine; by W. Wayt Gibbs; 8 Page(s)

Since about 1995, microprocessors have been outrunning the other parts of computer systems by ever increasing margins. The latest processors churn through instructions at up to 3.6 gigahertz (GHz); some operations, such as arithmetic, run at double that rate. But the wiring on the motherboard that connects the processor to its memory chips and other pieces of the system plods along at 1 GHz or less. So the brain of the machine spends as much as 75 percent of its time idle, waiting for instructions and data that are stuck in traffic.

"In the coming years, the imbalance between microprocessor performance and memory access will be driven to a crisis point," physicist Anthony F. J. Levi of the University of Southern California argued in a detailed analysis three years ago. He noted that the plastic material in printed circuit boards squelches high frequencies: for every 2-GHz increase in electrical signal bandwidth, signal strength falls 10-fold. As clock rates rise, so do power consumption, heat production and electromagnetic interference. Those are already three of the biggest headaches for system designers. And International Sematech, an industry consortium, forecasts that processor-to-peripheral links must accelerate by roughly 2 GHz every two years just to keep the bottleneck from tightening further.