Moore's Law is 50 years old

[alanb]

I've just read that Moore's law is now years old, it was first published in the April 1965 edition of Electronics Magazine.

My rough calculations show that in fifty years there are approximately 33 x 18 month periods. 2 to power of 33 gives 8,589,934,592. That's a big increase in chip power.

[BradC]

It must be getting close to the end of it's surprisingly long life though. Transistors can not get too much smaller.

And yet every time someone says something like that, it's exactly what happens. Life is never boring when you follow technology

[Richard Crowley]

It must be getting close to the end of it's surprisingly long life though. Transistors can not get too much smaller.

Yeah, we've been saying that for 30 years.  But the reasons for thinking that have changed over the years.  Many of them related to fabrication methods like lithography and how to image such small details. But they are using very tricky things like shorter wavelengths, step-and-repeat, reduction (vs enlargement or same-size masks) and interference lithography.

Now, we think we are approaching the limits of the mollecular structure of the silicon material, so perhaps a major sea-change in basic materials is imminent.  Although they continue to find ways of using and modifying silicon to keep the old gal alive.

[Mechanical Menace]

Now, we think we are approaching the limits of the mollecular structure of the silicon material, so perhaps a major sea-change in basic materials is imminent.  Although they continue to find ways of using and modifying silicon to keep the old gal alive.

A lot of it isn't so much down to the choice of materials as the quantum effects when building transistors out of a handful of atoms instead of relatively large clumps of atoms. 7 atoms seems to be the limit, less than that you get into the fuzzy realm of quantum computing. I personally can't see quantum computing being useful for general purpose work but I'd love to be proven wrong on the matter.

[sean0118]

We could try and split the atom but that could be slightly dangerous... 

[Tallie]

Well unless the silicon atom get smaller sometime soon they are pretty much stuck at that limit. Sure they can stack multiple layers but that has limits too.

Who's to say we'll still be using silicon?

[LabSpokane]

It's not a law. It was an observation/hypothesis that had to be modified to fit the data. 

The silicon industry labeled it's successive technologies according to Moore's "observation" even though what was making chips faster was not simply a matter of how many transistors were packed onto a piece of silicon.

"Moore's Law" is right up there with "IoT" on the list of terms that help me identify the speaker/writer as a moron. 

[alanb]

Here is a link to a reprint of the original article http://www.cs.utexas.edu/~fussell/courses/cs352h/papers/moore.pdf

It's not a law. It was an observation/hypothesis that had to be modified to fit the data. 

The silicon industry labeled it's successive technologies according to Moore's "observation" even though what was making chips faster was not simply a matter of how many transistors were packed onto a piece of silicon.

"Moore's Law" is right up there with "IoT" on the list of terms that help me identify the speaker/writer as a moron. 

I agree it's not a Law in the true sense however it's universally known as 'Moore's Law' not 'Moore's observation'

[suicidaleggroll]

It's not a law. It was an observation/hypothesis that had to be modified to fit the data. 

The silicon industry labeled it's successive technologies according to Moore's "observation" even though what was making chips faster was not simply a matter of how many transistors were packed onto a piece of silicon.

"Moore's Law" is right up there with "IoT" on the list of terms that help me identify the speaker/writer as a moron.

 
Of course it's not a law, just like Murphy's law isn't a law.  That doesn't mean that anybody who quotes it is a moron...

[Tallie]

Well unless the silicon atom get smaller sometime soon they are pretty much stuck at that limit. Sure they can stack multiple layers but that has limits too.

Who's to say we'll still be using silicon?

What did you have in mind as an alternative? Does it still have atoms?

Quantum computing is what I was hinting at...

[Richard Crowley]

It's not a law. It was an observation/hypothesis that had to be modified to fit the data. 

The silicon industry labeled it's successive technologies according to Moore's "observation" even though what was making chips faster was not simply a matter of how many transistors were packed onto a piece of silicon.

"Moore's Law" is right up there with "IoT" on the list of terms that help me identify the speaker/writer as a moron.

 
Of course it's not a law, just like Murphy's law isn't a law.  That doesn't mean that anybody who quotes it is a moron...

@LabSpokane is apparently just trying to be pedantic. It says more about him than it does about people who use the term.   

[LabSpokane]

It's not a law. It was an observation/hypothesis that had to be modified to fit the data. 

The silicon industry labeled it's successive technologies according to Moore's "observation" even though what was making chips faster was not simply a matter of how many transistors were packed onto a piece of silicon.

"Moore's Law" is right up there with "IoT" on the list of terms that help me identify the speaker/writer as a moron.

 
Of course it's not a law, just like Murphy's law isn't a law.  That doesn't mean that anybody who quotes it is a moron...

@LabSpokane is apparently just trying to be pedantic. It says more about him than it does about people who use the term.   

You're correct. It says that when you spout "Moore's Law says..." to me in a business meeting that I'm very likely going to show you the door shortly thereafter. 

It was a good observation in its day, only half of which is typically remembered properly, but it's time to call it what it truly is, and that is not a "law."  And it's particularly time to stop applying it to anything besides etching transistors on substrates during the period of time that it actually mattered.

[photon]

5nm cmos transistor gate length is predicted to be the end of Moore's Law. But that's a 2 dimensional law. Already, the silicon die are being stacked upon each other to make a 3 dimensional chip. Someone will have to invent a law for this.

[CatalinaWOW]

It may not be a law, but it has been an organizing principle for an industry for a long time.  Silicon wafer makers look at what feature sizes will be needed to fulfill this "law" over the next few years, and plan their product purity and planarity based on the "law".  Photolith mask makers and the various resist and other chemistry types do similar things.  Circuit designers develop the tools needed to map and plan such large component counts.  Testers develop their machines for the speed and probe resolutions require.  Product designers develop designs based on what will be possible in a given time window.

The "law" has been the symphony director making sure that all of the various technologies come on line at the right time.  Even banks and other funding sources are involved.  Just try to get a loan to develop equipment that is out of step with the "law".

[Richard Crowley]

You're correct. It says that when you spout "Moore's Law says..." to me in a business meeting that I'm very likely going to show you the door shortly thereafter. 

You must have me confused with someone else. I never said "Moore's Law says...".
For that matter, Dr. Moore disclaims the whole "Law" thing and rather agrees with you.  Which major semiconductor company did you co-found?

[LabSpokane]

For that matter, Dr. Moore disclaims the whole "Law" thing and rather agrees with you.  Which major semiconductor company did you co-found?

None.  And that's exactly my point.  Moore's observation/hypothesis/postulate/whatever should remain as applied to transistors on substrates.  It's been bastardized into some narcissistic, universal birthright that everything must become exponentially more powerful or cheaper because that's what's happened with transistors.  Silicon Valley took "Moore's Law" and tried to apply it to the liquid fuels industry with absolutely disastrous results. (Google: "Vinod Khosla")

The miniaturization of the transistor is a spectacular success story and those that did it are without doubt, giants, but "Moore's Law" is a unfair, wretched yardstick by which to measure anything else.  It has literally cost billions in just US taxpayer money on ridiculous projects, founded solely on the belief that the skill to miniaturize a transistor can be used to equal effect in any other endeavor. 

If there is one outcome of the anniversary of "Moore's Law" that would be most fitting, it would be for the media, the pundits, and Silicon Valley Kool-aid drinkers to put the observation back into its correct, complete and proper context.  i.e. It is time to remove the word "Law" from its title. 

*That* would be worth celebrating.

[helius]

A law is any equation that describes the world accurately. Moore's original article about doubling capacities does not print an actual equation (and so does not propose a law), but it uses graphs from which equations can be inferred. The second graph shows a trend of log₂ N_{components/die} ? t . This was amended later, to reflect that the doubling period had shifted to every two years, when semiconductor technology matured.

The first graph is actually more interesting, as it shows a trough-shaped function of cost vs. integration at each process point: the lowest cost per transistor is not at the greatest possible level of integration, but the one that optimizes device yield against packaging and interconnection costs. This graph foretells multi-chip modules among other things.

The original article (in Electronics 38.8, April 1965) has some other interesting predictions:
"There is no fundamental obstacle to achieving device yields of 100%." Well....

His observations about linear (analog) circuits being less amenable to miniaturization because capacitance and inductance cannot be shrunk as much as gain elements, and the coming of monolithic differential amplifiers, are all very astute. The paper includes a great number of predictions, which have all (with the exception of 100% yields) been borne out.

[Richard Crowley]

Even large wafers (300m) with very complex LSI die have achieved 100% yield.  It is easy to imagine that people who are making diodes and transistors and even op-amps, and SSI gates, etc. are achieving ~100% yield unless they are really sloppy.

[helius]

wafer size is not really related to yield, although there is proportionally less areal loss on the edges of a larger wafer. yield is a function of defect density and die area. of course, you only count defects large enough to affect device operation, so it is possible to make a chip with a very large feature size not be affected by defects. but the whole point of the Moore paper is about the highest economical density devices, which will push feature sizes down and be affected by defects!

it has even become common to design asics with redundant elements as a way to improve yields. if it were possible to produce these devices with 100% yield, then these measures would not be necessary. I think Moore was prescient in some ways, but there have been a large number of changes that could not be foreseen in 1965. Rapid wafer ATE, steppers, automated ball bonders and dicers, just didn't exist.

[eas]

I think Moore was prescient in some ways, but there have been a large number of changes that could not be foreseen in 1965. Rapid wafer ATE, steppers, automated ball bonders and dicers, just didn't exist.

To someone's earlier point though, the present existence of all those technologies was likely catalyzed by the initial correctness of Moore's astute observation. By describing an existing emergent phenomenon of a complex system, he helped reinforce it.

wafer size is not really related to yield, although there is proportionally less areal loss on the edges of a larger wafer. yield is a function of defect density and die area. of course, you only count defects large enough to affect device operation, so it is possible to make a chip with a very large feature size not be affected by defects.

I thought achieving uniformity of many (but not all) process steps was more difficult at larger wafer sizes?

[Richard Crowley]

I thought achieving uniformity of many (but not all) process steps was more difficult at larger wafer sizes?

Of course it is. Hundreds of billions of dollars were spend on developing tools and processes to produce decent yields for 300mm (12 inch) wafers.
And the cost to go to 450mm (18 inch) wafers is so great that it will take all the big competitors cooperating to bear the development costs. AFAIK that hasn't been accomplished yet.

[ukee1593]

This Kinda relates to Moores Law;

I was discussing on another forum the efficiency of CPUs,

So Its my understanding that even after 50 years of transistor refinement, 99.9% of the electrical power that goes into a CPU gets turned into heat during operation.  Some people take that to mean that the CPU is only 0.01% efficient.  I don't think this is a valid argument. 

My argument is that efficiency of a CPU is measured in number of Operations (FLOP or something) per Watt, and is only useful to compare different processors to each other.  We simply cannot say whether our processor is 100% efficient at turning electricity into Math as its not a valid comparison. 

Am I going insane?  Do processors send the majority of their electrical energy back as heat? 

Forum is here if you are interested in looking: https://forum.teksyndicate.com/t/tdp-heat-created-electricity-consumed-by-the-cpu/79081

BTW, I seem to remember in computer science that Moore's Law is used as an excuse to classify any algorithm which is less than exponential (so polynomial) complexity can be solved quickly ... because "computers will get fast enough to solve it quickly in the future".  Kinda a cop out, but it really proves the relevance of Moores law in the industry.  It isn't just an Intel marketing fad!!

[helius]

You do not "turn electricity into math." Energy is conserved, and the sum of all the energy in the system at the beginning is equal to all the energy at any future time. All (100%) electrical energy that is put into a CPU becomes heat, since the CPU is not a power conversion device: there is no other form of energy it is designed to produce. In some cases there may be a small component of acoustic or optical energy produced, but that is not desired and will be extremely small. The electrical energy in watt-seconds that is formed by the Vcc and the supply current, net any data lines in or out, turns into the joules of heat emitted from the chip. More generally, the third law of thermodynamics implies that the final state of all energy in the universe is as heat.

Computation itself is not an operation involving energy, and can happen with no energy being consumed (http://en.wikipedia.org/wiki/Reversible_computing). What requires most of the power in a practical computer is driving signals over wiring (both inside the CPU and between chips and I/O ports). Every real wire has parasitics (resistive, capacitive, inductive) that consume energy from the signal. This spent energy becomes heat, and if there is too much heat emitted in a small space, it must be removed actively, which uses even more power to drive fans or pumps. So energy efficiency in computer terms involves minimizing wire lengths, signal voltages, and cooling requirements. In the context of the Moore equation, shrinking the transistors on a chip reduces the voltage needed to switch them, and allows the wires between them to be shorter. When you can integrate more functions on a single chip you also eliminate circuit board wires.

[SeanB]

IIRC most of the power used by a CPU is used in clock generation, distribution and termination, easily the largest part of the die itself when added up. That is why the newer Intel processors moved to sections that are clockless, so that the only clocks are at input and output to latch the data in and out. IIRc nearly 60% of the total area of the die is used for the clock, easily equal to a massive power device.

The rest of the power is actually used in leakage current in gates, as the channels are so small now that off state leakage currents are going to be significant. It might only be tiny per gate, but have a million of more and it gets large.

[TimFox]

The first thing I learned in sophomore physics was that many physical equations have the solution X = A exp (+b t), but that those solutions are invalid since they have already, or will soon "blow up".
However, many situations seem to fit positive exponentials at the time that we are observing them.  The usual real solution is sometimes called the "logistic function"
f(x)  =  L / [ 1 + exp (-k (x - x0) ) ]
which is the usual S-shaped curve.  Initially, for x near -6, the curve is close to exponential, but it approaches a constant at large time.
This function is found in population studies, where the initial growth is fast, but there is a limit due to finite resources.  In this discussion, the limit is probably due to atomic dimensions.