Slashdot Mirror
Moore's Law Disputed
Kumiorava writes "Transistors can be packed to same chip two times more in every 18 months. This Moore's law has been repeated already over 30 years. Computers become faster, IT economy grows, but Moore's law doesn't apply. That has been proven by researcher Ilkka Tuomi. You can read the research from First Monday article The Lives and Death of Moore's Law." 'tho, to be fair, it seems to me that Moore's Law has lasted a lot longer then the throng of people who keep predicting its death.
It will stop, right? I mean, if the marathon record gets 10 minutes shorter every few years, for example, that doesn't necessarily mean that 100 years from now we'll be running a 20 minute marathon.
Aren't there limits to materials and stuff like that, or do we come up with Infinite Probability Drives, Dimensional Transfunctioners, Flux Capacitors, Heisenberg Compensators, Ludicrous Speeds....
Another factor is the great disparity between actual processing power (often measured in FLOPS etc) and the number of transistors on a chip. For a while, transistors numbers were doubling every 12 months, but computing power was only doubling every 24 months. Why? The need for pipelining and data management meant more and more of the chip had to be dedicated to pre- and post-processing of the actual calculation, along with intelligent caching and the related works of predictive streams.
An alternative approach has been to build specialized hardware to put all those transistors to use, at the expense of turning your general purpose computer into a very special purpose machine. This has been used, sometimes to great effect, in for example N-body calculations (GRAPE 1-6), yielding 50 or more TFlops of performance for the general computer cost of a 500 GFlop machine. It provides yet another example of the misappropriation of Moore's law.
The idea of the paper is to show that Moore's law can't be used to predict trends in economics.
So
a) "Moore's law" shows us the effect of demand vs. supply
b) It does not mean that the demand (or demanded quantity) would increase infinitely
c) You can not call it a law because the variations have been too big (first it was one year, then two, now 18 months) and as the formula is that of exponential growth, those variations mean huge differences at the number of transistors over a period of, say, five years.
In short, this article looks at the economics (as in macroeconomics) side of Moore's law. It doesn't claim that you couldn't pack more transistors or whatever on a microchip.
You could also claim that Moore's law might actually hinder economic development as Intel wants to obey the law. What results is that we are actually saying that "wow, Intel is keeping up with the R&D forecasts stated in their company strategy". Yipee.
Okay, a shitty explanation but please read the paper and look at the idea behind it before saying it's total bullshit.
Did anyone actually read the damn article?
It's about how the entire concept of Moore's Law is vague and has been applied to all sorts of other things exhibiting exponential growth, even though Moore was not referring to them. And specifically Moore never gave the time frame of "18 months." He said "1 year" one time, then later said "2 years." And if you look at the data, the transistor count of chips doubles roughly every 26 months, not 18. The point of the article is that Moore's Law is more of a hazy myth than anything else.
It looks like about 3 people so far, but some read it more carefully than others. Please everybody who is reading this: read this article article because it is very important. Again, though, even people who have read the article (or skimmed it) appear not to have gotten the full message. So Junks Jerzey writes:
It's much worse than that, actually. When he really pulls the gloves off and looks at the hard data over the entire 43-year history of the industry, he finds *no* simple doubling time for almost any measure of interest that has been claimed to be Moore's Law or any folk version of it. Even for transistor counts. What you can sometimes sort of show is iffy exponential fits to the data for 5-10 year periods. Strikingly, though, the doubling rates for several of the measures the author investigates have *slowed*. Improvements do keep on happening, but the pace of the improvement is not as consistent or rapid as you might have expected.
Now the big deal about this is simple. Anybody who tries to project that our problems will be solved when X doubles in Y months is really walking on thin ice. It is also important because chip technology has often been held up as some special and amazing business whose success should be inspirational to us all, since it improves so fast. Clearly, improvements in raw components have been rapid (although not as rapid as you might expect), but the Big Changes caused by technology are rarely tightly coupled to the speed of improvement in underlying technology. Hey, the *big* change of the last decade is that your grandma now probably has email. I'm not sure it makes sense to calculate how many transistors that took.
Babar