Linked VCSEL arrays point to parallel processing

Sunny Bains

A freespace optoelectronic system that can interconnect 64 processors has been built at the NEC Research Institute in Princeton, NJ. The system is based on a new computer architecture called the interconnectioncached network. Information is carried between multiprocessor boards using verticalcavity surfaceemitting laser arrays (VCSEL arrays) and passive optical channels (see photo). Within each board, data are routed by an electronic crossbar switch. Researchers say that the system could be implemented in realworld applications without dramatic improvements in optical technology: limited progress in the size and reliability of VCSEL arrays is all that is required.

According to researchers Eugen Schenfeld and Ian Redmond, the key to the new system lies in its approach to parallel computing. With Vipul Gupta, a doctoral candidate at Rutgers University (New Brunswick, NJ), Schenfeld looked at the various networks used to allow multiple processors to work on tasks simultaneously: the hypercube is one wellknown example. What they found was that, most of the time, each processor spent all of its time talking to perhaps two or three others; connections to all the remaining processors were not as important. Their conclusion was that, as long as the switching within each of these clusters of processors was very fast, the switching between clusters could be slower without slowing the computer down.

The next stage was to implement these conclusions in a physical system. Because of their speed, electronic crossbars were the clear choice for local switching. Though his background was in computer science and electronics, Schenfeld decided to use optics for the longer links. As a result, in 1992, he brought in Redmond, who had previously worked on opticalcomputing projects at HeriotWatt University (Edinburgh, UK), to research and design the optical system.

The result is a system that interconnects four processor boards, each representing 16 simulated processors, via four 8 ¥ 8 VCSEL arrays. Each processor is attached to four lasers on the array, each of which represents a connection to one of the four boards. The laser light travels through a passive system of microoptic arrays, bulk optics, and beamsplitter cubes until it arrives at the corresponding processor of one of the other boards. Which board is determined by which laser is on. For instance, if processor 14 on board 1 turns on its laser number 3, it will send a signal to processor 14 on board 3. If necessary, the signal can then be quickly routed to another processor on board 3 via the electronic crossbar.

According to Redmond they decided to use VCSELs rather than spatial light modulators because they are fast and avoid unnecessary power consumption. A spatial light modulator modulates an incoming beam, implying that the beam is there, drawing power, even if the modulator is going to turn it off again. With the NEC system, each laser is turned on only when it has information to transmit; therefore, power isn`t wasted. Despite the fact that current VCSEL arrays have performance too poor to be practical for realworld computers, Schenfeld and Redmond remain optimistic. Even modest progress in this technology, they say, could make their architecture practical outside the laboratory.

SUNNY BAINS is a technical writer/editor based in Boston, MA.

Click here to enlarge image

Simulated processor boards are linked through VCSEL arrays and a passive optical system. The VCSELs are stimulated from outside the system to provide processor output. After traveling to the input port of the next processor, light from the VCSELs is collected by fibers so that it can be detected and analyzed.

Sun Jan 01 00:00:00 CST 1995