CHAPTER FOUR – HIERARCHICAL MODEL OF COMPUTATION

Publisher Summary This chapter presents a simple model of computation for a graphics supercomputer. The chapter explains how and why Titan uses general techniques of concurrency. A graphics supercomputer is designed to take input data, to perform many computations, and to display the results as a high-quality graphics image. On small computers, all phases of the program execution are often handled by a single processor. This yields a simple system but does not result in very high performance. In Titan, concurrency by replication and specialization is used to increase the computation rate. Titan approaches the problem of providing specialized hardware for important functions with the philosophy of providing the minimum required hardware consistent with high-speed performance. The operating system phase requires a general-purpose integer processor unit. The computation phase usually involves floating-point arithmetic. For this reason, Titan has a sophisticated vector floating point unit with hardware support for floating-point multiplication, addition, and division. A further level of breakdown within instruction execution is possible because each reference to memory results in a cache or virtual memory miss that requires still more processing.