The following is a guest post by Dave Turek, vice president, IBM Deep Computing. IBM is well known for dominating the Top500 lists of supercomputers. Less well known is the Green500 List, which ranks supercomputers not only on feeds and speeds, but energy consumption. This year, IBM dominated that list with 17 out of the top 20 machines on the list. This blog post highlights the significance of this achievement for Smarter Planet.
Energy efficiency is quickly becoming one of the most important metrics of supercomputing value. Just a few short years ago high performance computing (HPC) clients were concerned primarily with performance, and the cost of performance. The conversation has shifted dramatically. HPC clients are now equally concerned about power consumption and cooling requirements. For good reason, the cost to power an HPC environment today is nearly as much as the hardware.
Fifty percent of the energy consumed in today’s average data center goes toward cooling the systems and preventing overheating. Overheating, in turn, leads to reduced reliability. In fact, Wu-Chun Feng of Virginia Tech, one of the founders of the Green500 List of energy-efficient supercomputers believes that for every 10 degree Celsius increase in temperature the system failure rate doubles.
Significance of the Green500 List
All of this makes the Green500 List of the world’s most energy efficient supercomputers even more compelling, since it clearly calls out which companies are taking supercomputing innovation to the next level. The Green500 List is published by Green500.org and provides a ranking of the most energy-efficient supercomputers in the world. The most current Green500 list published in July reported that 17 of the 20 most energy efficient supercomputers in the world are from IBM. The number one system on the Green500 list is a three way tie between IBM QPACE systems in Germany based on the IBM PowerXCell 8i processor. These top three IBM systems each produce more than 773 Mflops (millions of floating point operations per second) per watt of energy.
IBM developers made energy efficiency a core design principle for HPC systems long ago. For example, when IBM introduced the BlueGene/P supercomputer, it more than doubled the performance of its BlueGene/L predecessor, but only consumed a little more power. The engineering focus on performance per watt is what makes IBM supercomputers the most efficient and consequently, the most affordable. Clients who need to deploy petascale performance systems can save one million dollars a year in energy costs by using IBM’s BlueGene/P, because it is over 40 percent more energy efficient than comparable supercomputers.
IBM recently delivered a first-of-a-kind water-cooled supercomputer to the Swiss Federal Institute of Technology Zurich (ETH Zurich) that consumes up to 40 percent less energy than a comparable air-cooled machine and decreases the carbon footprint of the system by up to 85 percent. The supercomputer, called Aquasar, is made up of water-cooled IBM BladeCenter servers, achieves a performance of six teraflops and has an energy efficiency of about 450 megaflops per watt.
IBM and Arctur, one of Slovenia’s leading software developers, recently announced an agreement to build one of the most powerful supercomputers in the region. Arctur will allow mid-market companies to lease time on the IBM System x iDataPlex high performance cloud computer and help reduce product development time by up to 75 percent. The iDataPlex platform maximizes performance per watt with innovative cooling techniques such as the IBM Rear Door Heat eXchanger. A single iDataPlex 2U server packs five times the computer power of a typical server while consuming 40 percent less power.
As the HPC community pushes toward exascale computing (the performance of one million trillion calculations per second), the energy efficiency design challenges are much greater than the transition from teraflop to petaflop. In previous supercomputing performance generations, the transition was primarily centered on processor development. If current technology design was deployed in exascale computing, such a system would consume between 100 and 200 megawatts (the equivalent energy of a small power plant) and come with a prohibitive power bill of more than $100 million per year.
Most HPC developers agree that an exascale system should only consume about 20 megawatts, which means getting to a 1000-fold increase in performance with only a 10-fold increase in power consumption. IBM is well on its way in research and development to create the next generation of systems that can achieve an exaflop of performance with acceptable levels of power consumption. The formula is a combination of processors, systems level architecture and software.
The global benefits of supercomputing are extraordinary, from pharmaceutical and genetic research to assessing financial risk with pinpoint accuracy or modeling the effects of climate change over the course of a century. Yet all of this work hinges on the ability to balance supercomputing capability with energy efficiency. Now, more than ever, supercomputing performance and power consumption must be developed and implemented with an equal measure of importance.
Dave Turek is vice president of Deep Computing for IBM.