If supercomputers weren't so useful for predicting weather and climate change, the tree-huggers of the world would insist they be outlawed. But supercomputers are here to stay and - so it seems - to suck up a lot of juice.
Twice a year, the Top 500 list of supercomputing sites racks and stacks the biggest, baddest machines in the world, testing their sustained performance on a set of Fortran benchmarks called Linpack. The list is compiled by a gaggle of supercomputer experts - Erich Strohmaier and Horst Simon, computer scientists at Lawrence Berkeley National Laboratory, Jack Dongarra of the University of Tennessee, and Hans Meuer of the University of Manheim - and being on the list gives system vendors and supercomputer centers some serious bragging rights. Not to mention some marketing help.
With the 31st edition of the Top 500 listing announced in June, IBM's "Roadrunner" hybrid Opteron-Cell machine - created for the U.S. government's Los Alamos National Laboratory - was the first supercomputer in history to break the one petaflop barrier. That's 1 quadrillion floating point operations per second. It was also the first time that the systems' power consumption data was made available.
And so, in this energy-crazed 21st century, you can now peruse the Green 500 list, which sorts the Top 500 by flops per watt.
The listing is interesting, in that trends you wouldn't think about become clear. Among the top 10 machines in the original list, the average machine consumes 1.32 megawatts and has a power efficiency of 248 megaflops per watt. (Roadrunner, at 437 megaflops per watt, is bringing up the class average - bigtime). But across the 50 biggest machines on the list, average power consumption falls to 908 kilowatts and average power efficiency falls to 193 megaflops per watt. And across the whole list of 500 machines, average power consumption falls to 257 kilowatts and efficiency falls further to 122 megaflops per watt.
On the Green 500 sorting, blade servers based on the QS22 Cell Power chip are coming it at 488 megaflops per watt - but these fairly small machines are ranked towards the bottom of the Top 500 list in terms of power. At the other end of the scale is Roadrunner, which hit the one petaflop performance level while consuming 2.35 megawatts. That gives you a 437 megaflops per watt rating on the green scale.
Of the top 50 machines on the Green 500 ranking, a ridiculously large number of machines bear the IBM label, being either blade clusters using Cell or X64 chips or BlueGene PowerPC clusters in large or small sizes. Interestingly, the BlueGene design is showing pretty linear scalability in terms of megaflops per watt, which was part of IBM's design goal a decade ago when it started the project.
The most energy-efficient non-IBM machine on the Green 500 list is a Xeon server cluster built by Silicon Graphics for Total Fina's oil exploration. It's ranked at number 10 on the Top 500 list, but with a rating of 240 megaflops per watt, comes in at number 17 on the Green 500 ranking. Similar machines are ranked in the 40s on the list, and you have to go quite a bit father down on the green list to start seeing Fujitsu, Dell, SGI, and Appro show up, and by then, you are down into the 150 megaflops per watt efficiency range.
Interestingly, number 499 on the list is the Earth Simulator super created by NEC, a box that was at the top of the Top 500 for a number of runs. This machine is the largest vector supercomputer every built - at least that anyone will cop to outside of a black ops budget. It has a rating of 5.6 megaflops per watt. When you turn it on and the lights in the entire country dim, you can't hide its existence for long.
Earth Simulator is still ranked number 49 on the Top 500 list, with just under 36 teraflops of performance. It's just a pity that this massively parallel vector box consumes 6.4 megawatts of juice. And the absolute worst green box is the "Thunder" Itanium-Quadrics box at Lawrence Livermore National Laboratory, which is rated at 19.9 teraflops but burns 4.9 megawatts. That's a measly 4 megaflops per watt. ®