Each node will have either 8 or 16 gigabytes of memory, aggregating to 750 terabytes of memory across the entire system. Communications among the nodes will go over IBM 5D Torus interconnects, capable of 40 gigabits-per-second throughput.
For an operating system, the compute nodes will run a Compute Node Lightweight open-source scalable kernel, and the I/O nodes and the front-end and service nodes will run a modified version of Red Hat Enterprise Linux. The system will be mostly water-cooled and consume an average of 60 kilowatts per rack.
IBM did not reveal the price for Mira, though it did say Argonne had purchased it with funds from a $180 million grant.
While this system will be capable of 10 petaflops, the Blue Gene/Q architecture should be able to scale to 50 petaflops and perhaps more, Turek said.
IBM developed the Blue Gene architecture in 1993 as part of a US$100 million development effort in conjunction with LLNL. The research effort was aimed at building out a highly scalable yet energy-efficient architecture for large supercomputers.
Sign up for Computerworld eNewsletters.