Energy efficient HPC?
Energy efficient HPC seems to be the trend of the hour – everyone talks about strategies to reduce power consumption in high performance systems. BlusGene/P gets a lot of credit for being so power efficient (at least if you consider the common but obviously insane LINPACK metric). Let us do some simple calculations in order to pinpoint the truth behind the hype.
An average federal HPC center in Germany takes about 1-1.5 MW to run today. That’s the equivalent of roughly ten sports cars running 24/7, or maybe 1000 (European) standard cars under average utilization.
What’s the number of standard cars in just a single big city like Munich?
Another comparison: Assuming there are 30 million households in Germany running their (quite power efficient) 100W TV set for two hours a day. If they all left the TV off for just a single day, that would save 6 million kilowatt hours of electrical energy. The HLRB2 federal supercomputer at LRZ Munich could run for almost a full year from this single day of abstinence.
What would you rather spend your high-quality electrical energy on: Supplying redundant non-information (a.k.a. standard TV entertainment) to millions of people who don’t need it anyway or giving a couple of hundred scientists a powerful research tool?
I’m not saying that power saving strategies are no use in HPC. The sheer problem of getting the power needed to a small spot like an HPC center is quite a challenge. Heat production is slowly getting out of hand and needs to be considered. However, arguing that “green HPC” could make a significant contribution to environmental protection is nothing but a convenient device to get funding for some obscure research projects.