Legend has it that when ENIAC, the first digital computer, was switched on for the first time, it drew so much power from the Philadelphia electrical grid it caused a brownout. Since then, computers— and supercomputers in particular— have always been significant consumers of electrical power. For the next generation of supercomputers, operating at petaflop speeds and working on things like detailed climate models, power consumption could represent a significant limiting factor. As Electronics Design Strategy reports,
In an irony of this environmentally conscious era, the supercomputers used to study issues such as climate change themselves impose a significant carbon footprint—consuming megawatts of electricity both directly and for the elaborate cooling systems that are required to deal with the excessive heat they generate. Even so, scientists wishing to tackle leading-edge research need 100× to 1000× more computing throughput than today’s high-end systems can provide.
Scientists at UC-Berkeley and the National Energy Research Scientific Computing Center have proposed a new supercomputer design «using millions of low-power embedded microprocessors instead of conventional server processors,» which is projected to use a fraction of the power of previous supercomputers. As Berkeley Research News explains:
To develop a 1-km cloud model, scientists would need a supercomputer that is 1,000 times more powerful than what is available today, the researchers say. But building a supercomputer powerful enough to tackle this problem is a huge challenge. Historically, supercomputer makers build larger and more powerful systems by increasing the number of conventional microprocessors — usually the same kinds of microprocessors used to build personal computers…. [A] system capable of modeling clouds at a 1-km scale would cost about $1 billion. The system also would require 200 megawatts of electricity to operate, enough energy to power a small city of 100,000 residents.
The proposed Berkeley-Tensilica computer, in contrast, would «consume less than 4 megawatts of power and achieve a peak performance of 200 petaflops.» According to Electronics Design Strategy,
The joint effort will focus on massively parallel designs featuring large numbers of processor cores connected via optimized links…. Each core dissipates just a few hundred milliwatts while churning out billions of FLOPS, representing an order-of-magnitude improvement in FLOPS per watt over traditional desktop or server processor chips, according to Tensilica. A supercomputer harnessing millions of such cores, tightly integrated at the chip, board, and rack level, will achieve the exascale goal within a power budget of «a few megawatts.»