A month ago, IBM and the University of Illinois broke ground for the data center that will eventually house the Power7-based "Blue Waters" massively parallel supercomputer. The data center, it turns out, is as tricky to design as the processor and server that will be humming along inside it.
During a talk at the recent SC2008 supercomputing event in Austin, Texas, someone said it would not be long before power companies would be giving away supercomputers to any governmental agency or corporation that signed a long-term power contract. This was a joke. But it brings into sharp focus the fact that power and cooling are the main limitations on scalability for supercomputers. That's been the case since the first Cray-1 vector super came out in 1976.
Stephen Elbert, a researcher at the Pacific Northwest National Laboratory, explained that there are plenty of data centers out there burning 60 to 70 megawatts and that a few have already broken through the 100 megawatts barrier. "Beyond that, you have to be your own power company." Or, to do a deal with one, as Google has done with the Bonneville Power Administration in its The Dalles, Oregon data center.
The biggest problem facing data centers, according to Ed Seminaro, chief Unix hardware architect at IBM and one of the people responsible for the Blue Waters project, is the inefficiency of the data center itself. Many of the tricks that IBM will be using in the data center that houses Blue Waters will probably be considered by other computing facilities and server makers. The reason is simple: There is no other way to get more computing done than focus on the data center.
By Seminaro's calculations, in the typical HPC data center, a 300 watt device running in a data center requires 800 watts of input power. "The efficiency is just awful," he said. From that 800 watts of input power required for a 300 watt server, the data center's power conversion hardware (including uninterruptible power supplies and AC distribution units) consumes 160 watts; cooling towers, condensers, evaporators, and air movers eat 140 watts. The cooling equipment inside the device (fans and blowers, but sometimes pumps for water blocks and jackets) eats another 50 watts.
The power supplies, which step down the juice so it can be consumed by the server itself, eats another 150 watts of that 800 watts of input power. To idealize this, it takes 2.67 units of server power to actually run a single server. And on a multi-petaflops performance scale, this kind of inefficiency gets very costly, very fast.
That's why the University of Illinois and the Blue Waters design is set to become the standard bearer for Power-based HPC server performance in the coming years, much as Lawrence Livermore National Lab has been for a decade with IBM's ASCI Purple and BlueGene/L supers.
IBM has been vague about the Blue Waters design, which seeks to create a multi-petaflops machine using approximately 200,000 Power7 processor cores. In the Blue Waters announcement back in July, all that IBM has said is that Power7 chips will be implemented in a 45 nanometer process and will have multiple cores, meaning more than two. But Ross Mauri, general manager of IBM's Power Systems division, has separately confirmed to me that Power7 will be an eight-core processor.
(He did not say what kind of cores would be in there. It could be a mix of Power and other vector cores, much like the Cell Power chip used in game consoles and in supercomputers today. This chip has a single 64-bit Power core and eight vector-style coprocessors).
According to an early presentation of the National Center for Supercomputing Applications (NCSA) at the University of Illinois, the Blue Eaters machine is supposed to have more than 800 TB of main memory (with at least 32 GB of main memory per SMP node). The initial goal is for 100 Gb/sec of external bandwidth, eventually quadrupling that to 400 Gb/sec.