What Do Computer Architects Design Anyway?

Abstract

This paper is an attempt to provide a tutorial answer to the question: What do computer architects design anyway? The target readers are people concerned with automatic verification of computer hardware. The goal is to help this audience understand both the nature of the circuits which are being designed today and some of the factors that influence the design process. At the very least, automatic verification techniques must be mature enough to be useful in proving correctness of some sufficiently large component of contemporary designs before designers can be expected to make much use of the technology. Additionally, history has shown that mere functional correctness is an insufficient reason to adopt a new method. New methods must be able to build upon present practices and techniques. If a new method requires abandoning the bulk of present practice, it is seldom accepted. The paper presents a skeptical view of automatic verification systems. The domain, in practice, is both too imprecise and too complex to generate much confidence in the short-term future existence of useful verification systems. The hardware design domain covers a broad scope from theory to the thousands of people who design and build countless boards, chips, and systems every year. Until verification systems are capable of supporting a very broad piece of the overall spectrum, they will remain merely an interesting puzzle for the research community. On the other hand, when verifiers exist that are useful in broad practice, they will be a tremendous asset to the designer’s tool kit. The paper attempts to impart some aspects of both design and specification which should be part of the focus of the activity in automatic verification. Finally the paper presents a small catalog of real design fragments. It is hoped that these fragments will help verification efforts by being perverse cases which will serve as interesting small scale proof targets for new verification systems.