Spending Moore's dividend

Parallel Visions

The IT industry's remarkable progress over the past half century has been fueled by a simple fact that was codified in 1965 by Intel co-founder Gordon Moore in his eponymous law [1]. The number of transistors in an integrated circuit doubles every two years. This law is explained by constant advances in the manufacturing process, which over time have reduced the size of a single transistor to the nanometer scale and enabled ever-increasing speeds. The original Intel 8086 microprocessor sported 29,000 transistors and clocked at a then-blistering 5MHz; 28 years later, the 2.93GHz Intel Core 2 Duo boasted an impressive 291 million transistors, 10,034 times the original [2].

The first CPU to break the 3GHz barrier was an Intel Pentium 4 variant released in 2002 [3], but disregarding unreliable overclocking speeds or IBM's 5.5GHz liquid-cooled mainframe processor [4], the top clock speed has been hovering between 3 and 4GHz for the past 11 years. Power dissipation physics has put a limit on what can be done in terms of pure clock speed, yet Moore's law continues unabated, which poses the interesting question that James Larus concisely framed years ago as "spending Moore's dividend."

Transistors, Transistors, Everywhere

Early beneficiaries of transistor gains that could no longer be spent in traditional approaches have been cache sizes. This development has helped alleviate the yawning speed gap between a CPU and the RAM feeding it instructions and data to process. The Von Neumann bottleneck limits a CPU's actual usable speed, as instruction-fetch stalls and data-cache misses can force the CPU to wait for the slower subsystem to deliver more work to it. Bigger caches make

... comments powered by Disqus