Supercomputers and Energy Efficiency

We have all seen the various energy conserving techniques proliferate on our home and business personal computers, such as the power management policies for our hard drive, monitor and system. But what about supercomputers, are they immune from this same energy efficiency future we have been headed toward? Don’t supercomputers use an enormous amount of power and incur significant cooling costs? Well, there is growing evidence that energy efficiency of supercomputers is now a growing IT issue and gaining international attention.

The benchmark for high performance computers in use by national defense, engineering, science and financial analysis is the Linpack benchmark. The Linpack benchmark measures computing speed by performing tasks common to engineering tasks. The benchmark reports the result in millions of FLoating Point Operations Per Second or FLOPS. Floating point refers to the decimal point and can be thought of as a computer implementation of the scientific notation which allows for recording very large or very small numbers in a shorter method. For example, the speed of light can be expressed as 3.0E+08 meters/sec instead of 300,000,000 meters/sec. Up to now, much of the focus has been on the performance or price/performance of these supercomputers, however, this is starting to change as energy efficiency is becoming increasingly important to buyers due to the energy required for cooling and power consumption.

As current rule of thumb, 1 megawatt (MW) of power consumed by a supercomputer today requires another .7 MW of cooling to off-set the heat generated - and each megawatt of electric power costs approximately US$1 million per year. The Japanese Earth Simulator, used for climate modeling ironically, is a 35 Tera FLOP computer and cost US$10 million per year for powering and cooling (1). Additionally, it is important to note that reliability of these supercomputers is directly related to the increase in temperature and therefore we need better methods to reduce temperatures and not just continue to increase them as the trend has been. Temperatures have primarily increased due to increased density of transistors, increased number of processors and increased clock frequency – all of these same methods have contributed to the increasing supercomputer performance.

To drive down power usage and cooling requirements, a technique called Dynamic Voltage and Frequency Scaling is one method employed. Dynamic frequency scaling (also known as CPU throttling) is a technique in computer architecture where a processor is run at a less-than-maximum frequency in order to conserve power (2). Micro processor manufacturers, such as Intel and AMD have developed such products for commercial use.

In January 2006, the SPEC Power and Performance Committee began development of the first benchmark for evaluating energy efficiency. The members of this committee comprise many of the major vendors in supercomputing. Also, in 2006 a new web site was launched by the green500.org with the purpose to provide an energy efficiency ranking of all the world’s top 500 supercomputers. These efforts are good examples of the increasing awareness of the need to improve energy efficiency in the supercomputer industry.

(1)Wu-Chun Feng, Xizhou Feng, Rong Ge "Green Supercomputing Comes of Age", Green Computing 2008, page 18

(2)From Wikipedia, the free encyclopedia

By Jamie Johnson

Jamie Johnson is an associate partner at IBM and an active volunteer in the promotion of renewable energy and sustainability of our planet.  Jamie write articles and white papers for GreenITTools.com