The Tyranny of Transistors: What Counts about Software?

Our intuitions tell us that computer software is vastly better than it was 20 years ago -more diverse, more reliable, more powerful. However, the aggregate measures we use for software show only slow progress. Meanwhile, computer hardware basks in the glow of Moore's Law, and software seems to pale in comparison. While it is possible that software power does grows sluggishly in comparison to hardware, it seems more likely that hardware reaps the benefits of a better decision about what to count in describing progress over time. As we pursue value propositions in software, we should consider carefully the values that we pursue. A good start would be to switch from counting measures of our input effort to counting measures of the results of this effort 1. MOTIVATION Computer hardware has a reputation for performance and cost improvement at a rate unprecedented in the history of technology, with selected measures doubling every year and a half or two. The fact that the benefit can be reaped in either price of performance adds to the luster. Software, on the other hand, doesn't have a widely-quoted measure that shows how its power and value have increased over the years. Absent such a measure, we avoid the question and regale ourselves -and the world -with tales of cost and time overruns and other project failures. Absent a good measure, we sometimes use inappropriate measures -such as the long-standing perception that productivity increases at rates of only 4-6%, leading to doubling times measured in decades instead of years. The consequence is that software suffers relative to hardware, standing as the ne'er-do-well, if raffish, cousin to hardware's industrious and respectable leading man in the saga of information technology. Like the ne'er-do-well cousin, software doesn't get -or effectively claim -credit for its very real progress. Through its interest in Economics-Driven Software Engineering Research, the EDSER community proposes to change the objective of software development from the creation of functionality to the creation of value. It therefore behooves us to consider carefully what we mean by value. 2. THE MOORE'S LAW PROBLEM In 1965, Moore extrapolated from four data points that the number of transistors (switches) on an integrated circuit (chip) would double every two years. In fact, the doubling rate has been about 18 months for the subsequent three and a half decades. Furthermore, "Moore's Law" is widely interpreted not as a technical measure of switches per chip but as a broader, softer measure of computing system power that includes speed, storage capacity, and bandwidth, achieved through architecture and aggregation as well as chip technology. Costs have not grown in proportion the cost of the power growth is largely in design, which is principally the cost of the designers. Human costs grow roughly with inflation, doubling every decade or two rather than every year or two. Manufacturing costs have also seen only modest increases, but that bears on the number of computers in use, not the power of each one. So hardware has cultivated and earned a public persona of relentless growth in both power and price performance. The genius of Moore's Law as an aggregate description of hardware improvement is that it • counts something simple and understandable • counts the same thing over time • counts an output rather than an input • refuses to get embroiled in all the deficiencies of transistors as a measure Proceedings of the Fourth Workshop on Economics-Driven Software Engineering Research, IEEE Computer Society, 2002