It’s no secret that a main roadblock on the trail to exascale computing comes in the form of high power consumption. Throughout the history of computing, performance, and to a great extent, performance per watt, has grown exponentially in accordance to Moore’s law. However, there seems to be a growing divide between the power efficiency of smaller devices and the world’s fastest supercomputing clusters.
Jonathan Koomey of Technology Review wrote an article this week talking about the energy efficiency of computing devices. Specifically, he focused on the ratio of flops per kilowatt-hour (kWH), noting that the power required to perform a computational task is roughly halved every 18 months. Koomey points out that as processors have become increasingly power efficient, it has opened up whole new application areas, especially for battery-dependent mobile computing.
More than ever before, computers can simply do more with less. One of the most telling instances of this trend is seen in mobile and sensor-based devices. Koomey points to sensors that run on as little as 50 microwatts of energy and require no direct power source. Created by Joshua Smith from the University of Washington, these sensors gather energy from radio and television signals and relay weather information to indoor displays.
On the opposite side of the spectrum is Japan’s K Computer, which draws 12.7 megawatts of energy to deliver 10.5 petaflops. In this arena, performance is still the driver, with supercomputing users expecting roughly a 1000-fold increase in flops every decade. That’s about ten times the rate of increase that is naturally delivered by Moore’s Law. As a result, the annual cost of power for these high-end machines – about a million dollars per megawatt per year – is now a major limiting factor for new deployments.
Koomey points out that in theory the K Computer’s computational performance would be matched in the next 20 years by a device using less power than a typical toaster. Unfortunately, scientists and enterprises that need those flops today don’t have the luxury of time to wait.
While consumer-based devices are benefitting from the natural progression of energy efficiency delivered by the semiconductor industry, it’s simply not good enough to meet the needs of HPC users, especially for those at the top end. To reach exascale computing within reasonable power limits by the end of the decade, architectural innovation will be needed on top of what Moore’s Law will be able to deliver