figure 1: cycle of innovation in the computer industry.

requires

Increased processor performance

Slower programs

larger, more feature-rich software

High-level
languages and
programming
abstractions

larger
development
teams

figure 2: improvement in intel x86 processors; data from olukotum, ¹⁸ herb sutter, a principal architect at microsoft, and intel.

transistors (000) Clock speed (MHz)

10,000,000.00

1,000,000.00

100,000.00

10,000.00

1,000.00

100.00

10.00

1.00

0.10

1970

1975

1980

1985

1990

1995

2000

2005

2010

 

processor-performance growth and whether multicore processors can satisfy the same needs. In short, how did we use the benefits of Moore’s Law? Will parallelism continue the cycle of software innovation?

In 1965, Gordon Moore, a co-found-er of Intel, postulated that the number of transistors that could be fabricated on a semiconductor chip would double every year, a forecast he subsequently

17 reduced to every second year. ¹⁰ Amazingly, this prediction still holds. Each generation of transistor is smaller and switches at a faster speed, allowing clock speed (and computer performance) to increase at a similar rate. Moreover, abundant transistors enabled architects to improve processor design by implementing sophisticated microarchitectures. For convenience, I

call this combination of improvements in computers Moore’s Dividend. Figure 2 depicts the evolution of Intel’s x86 processors. The number of transistors in a processor increased at the rate predicted by Moore’s Law, doubling every 24 months while clock frequency grew at a slightly slower rate.

These hardware improvements increased software performance. Figure 3 charts the highest SPEC integer benchmark score reported each month for single-processor x86 systems. Over a decade, integer processor performance increased by 52 times its former level.

technology officer at Microsoft, memorably captured this wisdom with his four laws of software, following the premise that “software is a gas” due to its tendency to expand to fill the capacity of any computer (see the sidebar “Nathan Myhrvold’s Four Laws of Software”).

Support for this belief depends on the metric for the “volume” of software. Soon after Myhrvold published the “laws,” the rate of growth of lines of code (LoC) in Windows diverged dramatically from the Moore’s Law curve (see Figure 4). This makes sense intuitively; a software system might grow quickly in its early days, as basic functionality accrues, but exponential growth (such as the factor-of-four increase in lines of code between Windows 3. 1 and Windows 95 over three years) is difficult to sustain without a similar increase in developer headcount or remarkable— unprecedented—improvement in software productivity.

Software volume is also measured in other ways, including necessary machine resources (such as processor speed, memory size, and capacity). Figure 5 outlines the recommended resources suggested by Microsoft for several versions of Windows. With the exception of disk space (which has increased faster than Moore’s Law), the recommended configurations grew at roughly the same rate as Moore’s Law.

How could software’s resource requirements grow faster than its literal size (in terms of LoC)? Software changed and improved as computers became more capable. To most of the world, the real dividend of Moore’s Law, and the reason to buy new computers, was this improvement, which enabled software to do more tasks and do them better than before.

 

myhrvold’s Laws

A common belief among software developers is that software grows at least at the same rate as the platform on which it runs. Nathan Myhrvold, former chief

how Was it spent?

Determining where and how Moore’s Dividend was spent is difficult for a number of reasons. Software evolves over a long period, but no one systematically measures changing resource consumption. It is possible to compare released systems, but many aspects of a system or application evolve between releases and without close investigation, and it is difficult to attribute visible differences to a particular factor. Here, I present some experimental hypotheses that await further research