Updates to this story
IBM is set to continue the monumental growth in supercomputer speed by addressing one of the pivotal hurdles that is slowing speed, the interconnect bandwidth, with a new optical bus.
While processors and memory have increased in speed and capacity over the years, the connections between these parts have provided a bottleneck on speed, effectively slowing the rate at which supercomputers can advance. IBM, however, has managed to solve this problem with a new optical bus.
IBM's research teams have developed a new optical data-transfer system or optical bus, which it says is the first in the world and will help computers achieve speeds of 100 times what is currently available, allowing for the production of supercomputers that really deserve the name.
IBM has developed parallel optical devices that it calls “optochips”, which should address the problem of when a supercomputer needs over a million optical links, a nightmarish situation for electrical engineers. The optochips are transceivers which convert electrical signals to optical ones and then back again, which can achieve dozens and possibly hundreds of communication channels, improving connection bandwidth considerably.
The optochip allows transfer speeds of up to 15Gb/s per channel, a 50 percent increase on the previous record of 10Gb/s. With 16 channels available, that leads to an aggregate bandwidth of 240 Gb/s, while only consuming 2.2 watts of power. The optochip has a five times greater efficiency than current modules and can achieve 28Gb/s per square millimeter, making it the highest bandwidth density available, which is vital for large systems.
In 2011 IBM is to launch its Blue Waters supercomputer, which is expected to be 10 times faster than its Roadrunner machine, released in 2008. It will use over a million optical interconnects, compared to the roughly 40,000 optical interconnects on Roadrunner, highlighting the fact that the speed increase is intinsically linked to how many connections are available and how well they perform. In fact, the number of parallel optical models required for Blue Waters will match the current global annual production, making IBM the leading consumer in this field.
The IBM T. J. Watson Research Center discovered a roughly tenfold increase in processing speeds every four years. Based on this it forecast that supercomputers will reach speeds of 10 petaflops by 2012, 100 petaflops by 2016, and 1,000 petalops (or one exaflop) by 2020. IBM's Blue Waters will be somewhat ahead of schedule, managing 10 petaflops next year, with the possibility of being scaled to 16 petaflops soon after.