No abstract

In massively parallel computer systems for embedded real-time applications there are normally very high bandwidth demands on the interconnection network. Other important properties are time-deterministic latency and services to guarantee that deadlines are met. In this paper we analyze how these properties vary with the design parameters for a passive optical star network, specifically when used in a massively parallel radar signal processing system. The aggregated bandwidth and computational power of the radar system are approximately 45 Gbh and 100 GOPS, respectively. The analysis is focused on the medium access control protocol, called TD-TWDMA, for the time and wavelength multiplexed network. It is concluded that the proposed network is very well suited to this kind of signal-processing applications. We also present a new distributed slot-allocation algorithm with real-time properties.

In array radar signal processing applications, the processing demands range from tens of GFLOPS to several TFLOPS. To address this, as well as the, size and power dissipation issues, a special purpose "array signal processing" architecture is proposed. We argue that a combined MIMD-SIMD system can give flexibility, scalability, and programmability as well as high computing density. The MIMD system level, where SIMD modules are interconnected by a fiber-optic real-time network, provides the high level flexibility while the SIMD module level provides the compute density. In this paper we evaluate different design alternatives and show how the VEGA architecture was derived. By examining the applications and the algorithms used, the SIMD mesh processor is found be sufficient. However, the smaller the meshes are the better is the flexibility and efficiency. Then, based on prototype VLSI implementations and on instruction statistics, we find that a relatively large pipelined processing element maximises the performance per area. It is thereby concluded that the small SIMD mesh processor array with powerful processing elements is the best choice. These observations are further exploited in the design of the single-chip SIMD processor array to be included in the MIMD-style overall system. The system scales from 6.4 GFLOPS to several TFLOPS peak performance.

In