Playstation 3 Processor Powers The World's First Petaflop Supercomputer

Berkeley (CA) - The Top 500 supercomputing list has hit a new milestone. The most recent ranking, released earlier today, includes the first supercomputer that has broken the Petaflop barrier: The DOE’s Roadrunner system is 33 times faster than Japan’s Earth Simulator supercomputer that dominated the Top 500 ranking five years ago.

Two weeks ago we were wondering whether Roadrunner would make it into the summer Top 500 ranking and how strong the system will be. Today it is official, the system leads the most current Top 500 ranking by a far margin: The peak performance is rated a 1.38 PFlops, which is more than twice the performance of the BlueGene/L, also owned by the DOE, which is listed with 596 GFlops in second place. The Argonne National Laboratory’s BlueGene/P supercomputer came in third with 557 GFlops.

Roadrunner is installed the Department of Energy’s National Nuclear Security Administration (NNSA). It is the NNSA’s third IBM-built supercomputer, adding to two IBM-built supercomputers at the Lawrence Livermore National Laboratory: The original BlueGene/L (#2), completed in 2005, had 131,072 p5 processing cores and delivered a sustained performance of 280.6 TFlops. Last year, it was upgraded 212,992 cores to hit 596 GFlops. The 12,208 processor ASC Purple (#21) is estimated to provide a sustained performance of 75.8 TFlops. Roadrunner integrates 13,896 Opteron cores (6948 dual-core processors) as well as 12,960 Cell BE chips with a total of 116,640 PPE and SPE cores for a combined 130,536 cores. The Cell BE processors also represent the heart of Sony’s Playstation 3 game console.

The six fastest supercomputers in the world are installed in the U.S. with three of them using AMD’s Opteron processors. The fastest supercomputer using Intel processors and being installed outside the U.S. is a BlueGene/P system called Jugene, which is located at the Forschungszentrum Juelich in Germany. Jugene is rated at a peak performance of 223 GFlops.

The top system installed in the industry is owned by Total Exploration Production, a French oil company: The SGI Altix system based on Intel Xeon processor is rated at 123 TFlops. A surprise showing came from Tata Sons, the Indian manufacturer of the $2500 Nano car, which is confirmed to have the #8 supercomputer in the world with a peak performance of 173 GFlops.

According to the publishers of the ranking, 283 supercomputers in the list already use quad-core processors, while 203 systems are running on dual-core CPUs. Only eleven systems still use single core processors, and three systems use the Sony Playstation 3 processor with 9 cores. IBM remains the clear leader in the TOP 500 list in performance with 48% of installed total performance, compared to HP with 22.4%. Dell, SGI and Cray follow with 5.4%, 4.4% and 3.2%, respectively.

Looking back in history, the supercomputer ranked at #500 would have been listed at position 200 in the previous Top 500 just six months ago, which is an indication how rapidly the main field of the list is evolving. The turnover rate is a bit slower on the top, but still impressive. For example, Japan’s Earth Simulator, which dominated the ranking five years ago, has dropped to #49 with a peak performance of 35.8 TFlops. The system currently ranked at #100 would have been listed at #39 in the preceding list.

For the first time, the Top 500 list also includes a power rating. The top 10 supercomputers consume a combined 13.2 MWatts while providing a total performance of 4.1 PFlops. The most power-hungry system in the list is Roadrunner with a consumption of 2346 KWatts.

The authors of the list found that IBM QS22 Cell processor blades are currently the most power efficient supercomputing components (up to 488 MFlops/Watt), followed by IBM BlueGene/P systems, which deliver up to 376 MFlops/Watt. However, it was noted that Intel Harpertown quad-core blades are catching up "fast": The IBM BladeCenter HS21with low-power processors was rated at up to 265 MFlops/Watt and SGI Altix ICE 8200EX Xeon quad-core nodes at up to 240 MFlops/Watt.