Computing in a Parallel Universe
Multicore chips could bring about the biggest change in computing since the microprocessor
Why build chips with multiple processors? Why not just keep cranking up the speed of a single CPU? If the latter option were feasible, the chipmakers would be delighted to adopt it. They are turning to multicore systems only because the path to higher gigahertz seems to be blocked, at least for now.
The causes of this impasse lie in the peculiar physical and economic laws that govern the design of integrated circuits. The most celebrated of those laws is an economic miracle: As transistors or other components are made smaller and packed more densely on the surface of a silicon chip, the cost of producing the chip remains nearly constant. Thus the cost per transistor steadily declines; it's now measured in nanodollars. This extraordinary fact is the basis of Moore's Law, formulated in 1965 by Gordon E. Moore, one of the founders of Intel. Moore observed that the number of transistors on a state-of-the-art chip doubles every year or two.
Less famous than Moore's Law but equally important are several "scaling laws" first stated in 1974 by Robert H. Dennard and several colleagues at IBM. Dennard asked: When we reduce the size of a transistor, how should we adjust the other factors that control its operation, such as voltages and currents? And what effect will the changes have on performance? He found that voltage and current should be proportional to the linear dimensions of the device, which implies that power consumption (the product of voltage and current) will be proportional to the area of the transistor. This was an encouraging discovery; it meant that even as the number of devices on a chip increased, the total power density would remain constant.
Dennard's conclusion about performance was even more cheering. In digital circuits, transistors act essentially as on-off switches, and the crucial measure of their performance is the switching delay: the time it takes to go from the conducting to the nonconducting state or vice versa. The scaling laws show that delay is proportional to size, and so as circuits shrink they can be operated at higher speed.
Taken together, these findings suggest that our universe is an especially friendly place for making electronic computers. In other realms, the laws of nature seem designed to thwart us. Thermodynamics and quantum mechanics tell us what we can't hope to do; levers amplify either force or distance but not both. Everywhere we turn, there are limits and tradeoffs, and no free lunch. But Moore's Law and the Dennard scaling rules promise circuits that gain in both speed and capability, while cost and power consumption remain constant. From this happy circumstance comes the whole bonanza of modern microelectronics.
» Post Comment