Continued advances in the capability of Gallium Arsenide (GaAs)technology have finally drawn serious interest from computer system designers. The recent demonstration of very large scale integration (VLSI) laboratory designs incorporating very fast GaAs logic gates herald a significant role for GaAs technology in high-speed computer design:1 In this thesis we investigate design approaches to best exploit this promising technology in high-performance computer systems. We find significant differences between GaAs and Silicon technologies which are of relevance for computer design. The advantage that GaAs enjoys over Silicon in faster transistor switching speed is countered by a lower transistor count capability for GaAs integrated circuits. In addition, inter-chip signal propagation speeds in GaAs systems do not experience the same speedup exhibited by GaAs transistors; thus, GaAs designs are penalized more severely by inter-chip communication. The relatively low density of GaAs chips and the high cost of communication between them are significant obstacles to the full exploitation of the fast transistors of GaAs technology. A fast GaAs processor may be excessively underutilized unless special consideration is given to its information (instructions and data) requirements. Desirable GaAs system design approaches encourage low hardware resource requirements, and either minimize the processor’s need for off-chip information, maximize the rate of off-chip information transfer, or overlap off-chip information transfer with useful computation. We show the impact that these considerations have on the design of the instruction format, arithmetic unit, memory system, and compiler for a GaAs computer system. Through a simulation study utilizing a set of widely-used benchmark programs, we investigate several candidate instruction pipelines and candidate instruction formats in a GaAs environment. We demonstrate the clear performance advantage of an instruction pipeline based upon a pipelined memory system over a typical Silicon-like pipeline. We also show the performance advantage of packed instruction formats over typical Silicon instruction formats, and present a packed format which performs better than the experimental packed Stanford MIPS format.
Date of this Version