The Embedded Internet

Moore’s “Law” – which predicts a halving of the cost per transistor every 18 months – has remained remarkably accurate over the last four decades, yet there are those who claim that it is about to come to an end.

Proclaiming the demise of Moore’s Law is nothing new – it has been a perennial pastime of pundits. But they were wrong before, and they’re still wrong.

To write an obituary for Moore’s Law either indicates a lack of understanding of Gordon Moore’s original claims or a needlessly narrow view of its implications. A broader interpretation – predicting an exponential reduction in the cost of computing over time – is as robust as ever and shows no signs of abating.

In April 1965, Electronics Magazine published an article by Gordon Moore titled “Cramming more components onto Integrated Circuits.” Based on historical observations, Moore predicted that the cost per digital component would decrease by a factor of two every twelve months for integrated circuits that were designed for minimum cost. (That “minimum cost” condition is key: if the circuit isn’t dense enough, you pay for unused silicon. If the circuit is too dense, you pay a premium for exotic fabrication techniques.)

A few years later, the number of months per halving was revised to 18 months, but this exponential trend has continued unabated over the last four decades. The term “Moore’s Law” has become immensely popular, and has been expanded to refer to almost any exponential trend in technology.

But with its popularity comes its detractors. Every few years, naysayers predict the death of Moore’s Law, with thoughtful arguments as to why Moore’s Law cannot possibly continue. Fortunately, engineers appear to be oblivious to these arguments and continually revive Moore’s Law by “cramming more components onto integrated circuits”. So far, the engineers have prevailed over the naysayers, mostly by increasing silicon wafer sizes (from 25mm in 1965 to 300mm in 2001) and reducing feature size (from 2400 nanometers in 1980 to 45 nanometers in 2008).

The current deathwatch for Moore’s Law focuses on the fact that transistor gates are now only dozens of atoms wide, and to shrink them beyond the current dimensions will result in leaky, low-performance transistors.

But does this mean the end of Moore’s Law? Far from it.

First, there may be further advances in reducing feature size – we have already seen quantum-scale devices in the lab capable of capturing a single electron and digital switches built from carbon nanostructures.

Furthermore, recall that Moore’s Law predicts that it’s the cost of integrated circuit components that will shrink exponentially – not the size. Technologies that are larger but much cheaper – such as amorphous silicon devices printed at room temperature – could also extend its lifetime.

But most significantly, a strict interpretation Moore’s Law as a predictor of cost per component misses an important point: the cost of computation has fallen exponentially over the last six decades – longer than Moore’s Law itself – and there is no compelling evidence that this trend will end soon.

Consider this plot (view full size), derived from Hans Moravec’s excellent data set (started in 1997 and frequently updated) showing dollars per MIPS (million of instructions per second) over time. From 1945 to 1985, the cost of computation halved roughly every 17 months. This interval spans several technologies, from vacuum tubes, germanium transistors, discrete silicon transistors and ultimately integrated circuits. Since 1985, the cost of computation has continually halved every 10.3 months, and has spanned several computing form-factors, from mainframes to minicomputers to desktop PCs.

Based on historical evidence, it’s easy to predict continual exponential decreases in computing cost. More difficult is guessing which form of computing will lead the cost/performance race. Perhaps it will be cell phones or gaming machines, with low prices driven by sheer volume. (In the year 2000, the Sony PlayStation II offered far better MIPS per dollar than anything else on the market at the time.) Or perhaps it will be large purpose-built machines, modern day equivalents to Big Blue or Whirlwind. Massively multicore processors are starting to come available, which offer high chip yields with relatively small processor chips – these may continue to drive the cost curve down.

What does this imply? While it is true that our current technologies for integrated circuits may run up against physical limits in the near future, history has shown that a new technology will arrive in time to replace it. And regardless of what technology comes to the fore, a design engineer can still safely bet on long-term, exponential decreases in the cost of computation.

---

copyright © 2008 nbt ventures, all rights reserved

---

share the love: | | | | | |