Light Speed Computers

Laser chips could power petaflop computers

Laser communications chips capable of pumping data through the veins of gargantuan "petaflop" supercomputers have been demonstrated by NEC in Japan.

The communications chips can transfer information through optical fibres at a blistering 25 gigabits per second (a gigabit is a billion bits). This is a record for such components, according to NEC, and is many times faster that the purely electronic interconnects used in today's supercomputers.

Communications chips can convert electronic signals into optical ones. Using optical fibres to relay data between the chips is what may give this type of supercomputer the edge over previous ones using processors connected electronically.

NEC used a type of semiconducting laser diode called a Vertical-Cavity Surface Emitting Laser (VCSEL) which generates laser pulses in response to an electrical current. Researchers at the company created more efficient VCSEL devices by making the diodes from a blend of gallium arsenide and indium gallium arsenide - they used indium instead of the more conventional aluminium. This made it possible to transfer laser pulses more rapidly through optical fibre.

The new VCSEL chips could be used to make supercomputers of unprecedented power by routing data more efficiently between thousands of individual computer processors. NEC believes the chips could prove crucial to the development of the first petaflop class supercomputer - a machine capable of carrying out a thousand trillion mathematical calculations every second.

"Petaflop-class performance can be achieved in the next-generation supercomputer installed with the new VCSEL, in about 2010," Takahiro Nakamura from NEC's System Devices Research Laboratories told New Scientist.
Off-the-shelf

Such an achievement might enable NEC to regain the supercomputing crown that it held between 2002 and 2004 with the Earth Simulator - a supercomputer installed at the Japanese Agency for Marine-Earth Science and Technology in Yokohama, Japan. This is because efficiency with which purely electronic chips share data is a crucial bottleneck in supercomputer design. Most of today's supercomputers operate at a maximum speed of few teraflops (trillions of operations per second).

Many supercomputers are essentially made by linking up thousands of off-the-shelf computer processors. However, the current number one, an IBM machine called BlueGene at Lawrence Livermore National Laboratory in California, US, is made from customised components and is capable of a fearsome 360 teraflops.

While external experts agree that VCSEL chips could be used to construct formidable supercomputers, they say the cost of such components will also be crucial. "Raw bandwidth alone is not necessarily the most pressing issue for petascale computing," says John Shalf at Lawrence Berkeley National Laboratory in California, US. "The question is whether we can afford such components."

Furthermore, although VCSEL chips promise to be cheaper than comparable optical technologies - such as indium phosphide lasers - Shalf says a cheaper solution could be to combine several electronic connection channels in a single data "pipe". Another approach may be to send several optical signals through the same cable, a technique known as wavelength division multiplexing.
Unprecedented complexity

"The ability to go to 25 gigabits per second using VCSELs provides some opportunities for more cost-effective components, but that remains to be seen," Shalf told New Scientist. "You can be assured the market will provide the answer when these things become real products."

Horst Simon, another supercomputing expert at Lawrence Berkeley, adds that other issues will affect the development of the next generation of supercomputers. "Building a petascale system with a useful amount of memory - say at least 200 terabytes - and then powering and cooling this system will be the bigger challenges," he says.

Regardless of the challenges ahead, NEC is confident that petaflop supercomputers will be able to carry out experiments of unprecedented complexity. "It will be able to carry out entire simulation of the human body from genes and cell level to the organs and even the entire body," Nakamura adds. "Complex and detailed simulation of the behaviour of nano materials, from elementary particle to device level, is planned."