The technology of optical signal transmission has consolidated its position in the field of signal transmission, since signals can now be transferred at high speeds over a long distance. However, the application of this technology to computers is still undeveloped, and this is therefore one of the goals of the optical category in the RWC project.
Together with the RWC Tsukuba Research Center, we are carrying out research on application of optical technology to the vast networks involved in massively parallel computers, under the theme of "Research on Optical Interconnection Technology for Massively Parallel Computers."
Various attempts have been made to apply optical technology to computers, and these can be divided into two main categories: optical information processing and optical interconnection. In optical information processing, the operations are carried out by light itself, using technologies such as holography, and it is expected that this will help achieve high-speed parallel processing using optical signal.
Optical interconnection, on the other hand, is an attempt to implement the information transfer element using optical technology, since this is the facility best suited to the use of optics, leaving more complex operations to Si-LSI. This capacity would provide computers with optical data buses that would be capable of transmitting data at speeds as high as gigabits per second (Gbps).
Optical interconnections are subclassified into "interunit," "interboard", and "interchip" connections, depending on the connection distance. Various proposals have been made in RWC project with regard to methods of incorporating surface-emitting lasers and optoelectronic integrated circuits (OEICs) onto a silicon substrate. Our research focuses on implementing optical interconnections using optical fibers to connect boards inside massively parallel computers, in the "interunit" and "interboard" categories, as shown in Fig. 1.
At Hitachi, research and development are being carried out in relation to optical interconnection modules in order to implement interboard interconnection. This module has 12 lasers and 12 photodetectors monolithically integrated and connected with 12-port ribbon fibers. This module features 12 channels capable of high-speed transmission of over 250 megabits per second (Mbps) per channel, allowing large-capacity signals totaling up to 3 Gbps to be transferred over 100m in bit-parallel form with low skew (i.e., a small delta in the arrival times of parallel-expanded bit signals).
Our main effort is concerned with the mounting of this module on the massively parallel computer (RWC-1) currently being developed by the RWC Tsukuba Research Center, with the goal being to implement the world's first massively parallel computers connected via optical signals.
While bit-parallel signal transmission can eliminate the process overheads involved in MUX/DEMUX (multiplexor/demultiplexor) procedures to and from bit-serial form, there is a new problem of "skew," a difference of the arrival times of synchronized bit-parallel signals. The current requirement for interboard connection is to transfer 36-bit synchronizing signals operating with clocks at over 50 MHz. We are attempting to operate four interconnection modules in sync in order to transfer 36 bits directly without having to use MUX/DEMUX. This requires a high degree of homogeneity between the characteristics of the lasers integrated in the arrays, as well as a technique for controlling and suppressing various causes of skew, such as fluctuation in drive circuit characteristics and wire length in multilayer boards.
We believe that the characteristics of optical technology to carry out long-distance and high-speed signal transmission will be clearly demonstrated in bit-parallel communication in computers which requires strict skew control by solving the above problems and realizing high-density implementation of computing elements and optical transmitters and receivers on the single board. In addition, the use of lightweight optical fibers facilitates unit assembly and may give a solution to the load caused when driving long-distance electrical signal wiring using LSI and a bottleneck by increasing the numbers of pins. It can therefore be expected that the technology will have an influence on computer architecture itself.
This year, in parallel with experiments in the mounting of optical components on actual computer boards, we also started research on material devices for high-efficiency lasers as an essential technological component of the higher-speed and higher-efficiency optical interconnection technology that will be required in the future.
As our work progresses, we hope to contribute to the establishment of flexible information processing systems, the goal of the RWC project, with our research demonstrating the practicability of higher-speed and larger throughput optical interconnections.