With a total of 116,640 processor cores, including AMD Opteron processors as well as 12,960 IBM Cell processors similar to those created for the PlayStation 3, IBM has broken the one petaflop mark with the new Supercomputer commissioned by the American military administration. The $133 million machine*, called Roadrunner, will be used to solve complex issues related to nuclear weapons, and to address problems like the climate change.
While a remarkable achievement, Brough Turner pointed out in his blog a few weeks ago that Google might have surpassed the petaflop mark unnoticed. The Google Cluster is built on the principle of extensive parallelization with fault-tolerance managed by software. The index is partitioned over several servers, so that a single search query is processed in parallel by many processors, minimizing the overall response time for a search. A paper describing Google Cluster Architecture (dated 2003, but worth reading), underlines how Google selects commodity class PCs, as they provide the best performace/cost ratio, without the need of expensive hardware reliability as fault-tolerance is handled by software. Google is estimated to currently have more than 500.000 servers distributed across their data centers. Assuming a performance of 14.7 gigaflops for an AMD Athlon X2 4600 processor (a good performance/cost ratio two years ago), Google platform must be handling today more than 7 petaflops!!**
At a $500 cost per PC, Google would have spent $250 million. Still cheap comparable to IBM price for one seventh of the capacity.
Whether record or not, IBM has taken 11 years to improve their supercomputer performance a thousandfold. As the performance growth is not linear but exponential, the next thousandfold improvement, taking us to exaflops, should be expected sooner than 8 years from now. Singularity might not be that far after all.
Reference: NYT’s Military Supercomputer Sets Record
* less than $140 million Real Madrid is claimed to be ready to pay for Cristiano Ronaldo
** I have used a less powerful processor than Brough, in the assumption that a Quad Core in end of 2006 were less cost effective than an AMD dual core.