Has Moore's Law Stopped?

[DWW]

IIRC I always thought Moore's law was that every 18-20 months transistor density was to double, but more recently I heard that people claim that every 18 months CPU speed is suppose to double.

If that is the case, why hasn't it held out? 36 months ago 1GHz were available from both Intel and AMD (measured in their own performance too, not just their clock speed), today 4GHz should be available, but instead we top out at 3.0 and ~2.0 GHz for Intel and AMD respectively. What gives?

 

[Lord Evermore]

The "law" has been distorted and reworded a lot.

http://www.arstechnica.com/paedia/m/moore/moore-1.html#part1

There's a good explanation of it. It really doesn't apply to CPU power specifically at all.

The final quote on the page can be broken down this way (using example dates for illustration only):

January 2000: 50 million transistors on a chip is the most cost effective; putting more transistor in makes the cost less efficient, because the die size increases resulting in lower yields and the density can only be so high; putting fewer transistors in results in lower performance and features, and wastes die space.

January 2001: 100 million transistors per chip is the most cost effective; the process has been shrunk, so the die size may be smaller or only slightly larger; higher density results in more performance per unit of die area, and reduced power consumption.

Intel has a graph showing their scales: http://www.intel.com/research/silicon/mooreslaw.htm

Note that the law does not specify that performance will actually double. This can't be a guarantee because clock speed variations make all the difference. We know that on a clock for clock basis, the P4 can't beat a P3, even with double the transistors. But a 3GHz P4 has the same number of transistors as a 2GHz P4 (assuming equal cache and all that). And a 700MHz P3 has the same number as a 1.3GHz P3. So you can't really use doubling of performance as a measurement, nor actual transistor count on a single CPU type.

 

[Wingznut]

Gordon Moore's original paper (.pdf)

And yes, it's about transistor density... Not speed.

 

[RyanM]

Originally posted by: Wingznut
Gordon Moore's original paper (.pdf)

And yes, it's about transistor density... Not speed.

Regardless of Moore's original word for word statement, he clarified in later interviews that by transistor count, he meant speed.

But he didn't mean speed as in MHz, either, he meant the operations per second. A more accurate metric than that is Millions of operations per second per dollar - Or MIPs/$.

And if you look at the number of MIPs that CPUs have been able to do over the last 30 years, roughly every 18 months, the number of MIPs/$ has doubled. Period. So although what people really talk about isn't the actual Moore's law, it's what was implied in his original statement, and the spirit is certainly there.