BAD HARDWARE WEEK

The Blue Gene/Q system is the third generation in the IBM Blue Gene line of massively parallel, energy efficient supercomputers. The Blue Gene/Q architecture scales to tens of Petaflops. The network and the highly parallel message unit, which provides the functionality of a network interface card, are integrated onto the same chip as the processors and cache memory and occupy only 8% of the chip's area, including IO cells. The chip has 11 ports; each port can transmit data at 2 GB/s and simultaneously receive at 2 GB/s for a total bandwidth of 44 GB/s. The network consists of a five dimensional compute node torus with a subset of compute nodes using the 11th port to connect to I/O nodes. The five dimensional torus provides both excellent nearest neighbor and bisection bandwidth; a Blue Gene/Q machine has approximately 46 times the bisection bandwidth than that of a first generation Blue Gene/L machine with the same number of nodes. It can be partitioned into non-interfering sub-machines. A single network supports point-to-point and collective traffic such as MPI all reduces at near link bandwidth. The collectives can be over an entire partition or any rectangular subset of a partition. The network provides bit reproducible, single pass floating point collectives. The message unit has multiple injection and multiple reception engines that parallelize the injection and reception of messages, permitting full utilization of all the links on the torus.
The 16-cores on a die are designed to deliver 204.8 gigaflops at a power draw of 55 watts.

BAD HARDWARE WEEK: In ten years there could be supercomputer networks using 17 degrees of torus, but data centers are trying to converge on one link, they are trying to be as flat as possible," said Torsten Hoefler. And Sequoia processor offers 4Gflops per watt. All the others supercomputer manufacturers deliver barely more than 1Gflops per watt. Most of them deep below that limit.