A patent filing details IBM's plans to take the building blocks of the Sequoia supercomputer to a performance level of 100 PFlop/s.
According to the document, which stretches over 649 pages, IBM describes a BlueGene/Q system that is based on 524,288 processing nodes with 16-core PowerPC A2 processors that are able to handle 64 threads each. The system would include almost 8.4 million processing cores that are organized in 512 racks. The targeted performance is a peak of 107 PFlop/s.
The BlueGene/Q Sequoia supercomputer IBM is currently building for the Lawrence Livermore National Laboratory will use about 1.6 million A2 processing cores in 96 racks. The patent claims that each processing node will consume about 30 watts of power, which puts the 107 PFlop/s system at only 15.7 MW. That is rather impressive for a system with more than 8 million CPU cores.
The patent suggests that IBM has big plans with BlueGene/Q and especially its 5D torus network that connects the computing nodes among each other. There was no information when such a 100+ PFlop/s system could become reality. Sequoia is scheduled to go online in 2012 with a peak performance of about 20 PFlop/s.