Sequoia supercomputer takes the lead with 16 petaflops
In the top 500 of the world’s fastest supercomputers, the IBM Sequoia in the Lawrence Livermore lab has taken the lead. The calculation sample, which will be used for nuclear simulations, manages to achieve 16.32 petaflops.
The Sequoia, which was first presented by IBM in 2010, consists of 96 racks. The racks contain a total of 98,304 nodes, which in total account for 1.6 million hpc cores clocked at 1.6GHz. The nodes communicate via 5D Torus, a proprietary fiber optic technology from IBM, with a bandwidth of 40Gbps.
Traditionally, the Linpack benchmark is used to determine the computing power of the Sequoia supercomputer. After a 23-hour test run, the computing power came out at 16.32 petaflops. In theory, the system can even peak at 20 petaflops, but the Sequoia ran with an efficiency of about 81 percent during the benchmark round. With a score of 16 petaflops, the Sequoia passes Fujitsu’s Japanese K Computer in the semi-annual top 500 list. This supercomputer has 705,024 cores and manages to achieve 10.5 petaflops.
With a power consumption of 7890kW, the system is significantly more efficient than the K Computer, which consumes 12,659kW. According to IBM, this is made possible by an innovative water cooling system; along and over each node are thin copper tubes in which water is pumped around. Traditional air cooling would consume much more energy than water cooling used by IBM.
The Sequoia, which was purchased by the Department of Energy and would have cost $200 million, will mainly be used for running nuclear simulations. Through these simulations, the US government wants to be able to determine, among other things, how long the existing nuclear arsenal can still be deployed.
European supercomputers have improved their position in the updated top 500 list. The most powerful is the SuperMUC, an IBM iDataplex system based in Germany that manages to hit 2.8 petaflops. The Italian Cineca system is in seventh place with 1.7 petaflops, while the German calculation monster JuQUEEN BlueGene/Q, with 1.4 petaflops, is good for eighth place. Furthermore, within the top 10 we find the French CEA/TGCC-Genci math monster, also with 1.4 petaflops.