Slashdot Mirror
Andy Grove Says End Of Moore's Law At Hand
Jack William Bell writes "Intel chief Andy Grove says Moore's Law has reached its limit. Pointing to current leaks in modern chips, he says -- "Current is becoming a major factor and a limiter on how complex we can build chips," said Grove. He said the company' engineers "just can't get rid of" power leakage. -- But, of course, this only applies to semiconductor chips, there is no guarantee that some other technology will not take over and continue the march of smaller, cheaper and faster processors. I remember people saying stuff like this years ago before MOSFET." Update: 12/11 22:01 GMT by T : Correction: the text above originally mangled Andy Grove's name as "Andy Moore."
I wonder if there's a similar law representing toxicity to the environment of semiconductor manufacturing techniques.
I'm curious what kind of results the experimentation in superconductivity and semi-conductors will yield. They sound kind of mutually exlusive. But we may yet see Moore's Law revived and revised...
Course, that's probably 15 years away...
Seeing as he is a big part of a major CPU firm Intel, is he being short-sighted (which I doubt) or is he trying to brace the market for a slowdown in CPU clock speed?
It might help the company if expectations for new CPUs aren't higher than what they can produce.
Personally, my vote goes for optical CPUs as the wave of the future. Larger than curent CPUs might not be a problem if they don't put off much heat.
Moderation: Put your hand inside the puppet head!
I've always had issues with calling Moore's Law a "Law". Nobody has conclusively proven it. It should instead be called "Moore's Hypothesis" or "Moore's Theorem" if you're more optimistic...
A couple of things:
- Grove said basically the same thing you said- if better insulators or other technologies aren't developed, Moore's Law could become "redundant" in 10 years.
- That said, there are other ways to increase chip performance other than increasing transistor density according to Moore's law. Grove cites a few of them in that article (more efficient transistors, multiple cores, etc). So you will still be able to play the latest Quake in 10 years.
"The defense of freedom requires the advance of freedom" - George W Bush
Of course, I think something else will pop up (like the aforementioned optoelectronic switch, perhaps), since companies are resourceful folks. Academia is good about researching ways to reduce current leakage, and my prof says high-K dielectric insulators are a good way to reduce leakage through the gate. Whatever...something will come up.
My point is that the situation now is a lot more physically complex than that of, say, 1989 or something, where the limitation was "we can't go past 100 MHz because we haven't thought of a way to do it!" Now it's more "we can't go past [whatever]Ghz because of goddamn physics!"
By the way, anyone else think Gordon Moore gets a little too much by having a "law" named after him? I mean, sheesh...all he did was draw a freakin' best-fit curve on a plot of easily-found data. And on top of that, Moore's Law isn't a law at all...it's a statistic.
or am I wrong?
So we're running out of ways to pack more and more transistors into a device. There's still a ton of room to improve the layout of those transistors, the world is full of whines about x86 architecture.
This doesnt mean 'computers are as good as they're going to get', it just means the fabrication plants are as good as they're going to get.
I don't need no instructions to know how to rock!!!!
How many times do we have to hear people put their foot in their mouth? I would have thought Intel would've known better!
... is it good for?
But what
- Engineer at the Advanced Computing Systems Division of IBM, 1968, commenting on the microchip.
I think there is a world market for maybe five computers.
- Thomas Watson, chairman of IBM, 1943.
What can be more palpably absurd than the prospect held out of locomotives traveling twice as fast as stagecoaches?
- The Quarterly Review, England (March 1825)
The abolishment of pain in surgery is a chimera. It is absurd to go on seeking it. . . . Knife and pain are two words in surgery that must forever be associated in the consciousness of the patient.
- Dr. Alfred Velpeau (1839) French surgeon
Men might as well project a voyage to the Moon as attempt to employ steam navigation against the stormy North Atlantic Ocean.
- Dr. Dionysus Lardner (1838) Professor of Natural Philosophy and Astronomy, University College, London
The foolish idea of shooting at the moon is an example of the absurd length to which vicious specialization will carry scientists working in thought-tight compartments.
- A.W. Bickerton (1926) Professor of Physics and Chemistry, Canterbury College, New Zealand
[W]hen the Paris Exhibition closes electric light will close with it and no more be heard of.
- Erasmus Wilson (1878) Professor at Oxford University
Well informed people know it is impossible to transmit the voice over wires and that were it possible to do so, the thing would be of no practical value.
- Editorial in the Boston Post (1865)
That the automobile has practically reached the limit of its development is suggested by the fact that during the past year no improvements of a radical nature have been introduced.
- Scientific American, Jan. 2, 1909
Heavier-than-air flying machines are impossible.
- Lord Kelvin, ca. 1895, British mathematician and physicist
Radio has no future
- Lord Kelvin, ca. 1897.
While theoretically and technically television may be feasible, commercially and financially I consider it an impossibility, a development of which we need waste little time dreaming.
- Lee DeForest, 1926 (American radio pioneer)
There is not the slightest indication that [nuclear energy] will ever be obtainable. It would mean that the atom would have to be shattered at will.
- Albert Einstein, 1932.
Where a calculator on the ENIAC is equipped with 19,000 vacuum tubes and weighs 30 tons, computers in the future may have only 1,000 vacuum tubes and perhaps only weigh 1.5 tons.
- Popular Mechanics, March 1949.
(Try the laptop version!)
There is no need for any individual to have a computer in their home.
- Ken Olson, 1977, President, Digital Equipment Corp.
I have traveled the length and breadth of this country and talked with the best people, and I can assure you that data processing is a fad that won't lastout the year.
- The editor in charge of business books for Prentice Hall, 1957.
[Quotes from this page.]
If you celebrate Xmas, befriend me (538
Hmmm- I didn't see anywhere in the article where Grove said it is "unsolvable". Lets read what the article actually said:
He said the company' engineers "just can't get rid of" power leakage.
Sounds to me like he is just saying Intel hasn't solved it yet (but neither has anybody else).
"The defense of freedom requires the advance of freedom" - George W Bush
You haven't convinced me that the situation circa 2003 is any different than that c. 1989. As then, physics is placing limits on the performance of current design processors.
I think it is exceedingly likely that there will be advances in materials science and manufacturing that will prolong the validity of Moore's Law. It continues to be feasible to decrease core voltages, and newer heat-removal technologies and better dielectrics are showing promise. Even if each avenue provides only a linear reduction in dissipation, or a linear increase in our ability to deal with it, the end result is that the synergy allows us to eke out a few more years of exponential growth.
Lather, rinse, repeat.
Soon after AMD tells the world that they will no longer be completive with Intel, Intel comes out and says processors wont be getting faster as they used to.
Yeah NO SHIT!
I attended a talk some 1.5 years ago by guy from Philips NatLab (home of the CD), which was called "Mo(o)re or less?". Although the talk was extremely boring and i forgot the final conclusions i do remember some potential showstoppers he listed:
-Of course the ultimate limit of a 1 atom transistor, can't remember the date this would occur
-Limited speed of signals acros the chip: If the clock frequency gets much larger a signal would require several buffer stages to reach the other side.
-Capacity of wires gets more important: the interconnects don't scale at the same pace as the transistors. Their finite capacity limits clock speeds
Some non-technical reasons:
-Increasing costs of new fabrication processes: each new increment is more expensive.
-Limited manpower to design circuits with more and more transistors. This probably means that a larger area of the chips will consist of 'dumb' circuits like cache.
karma police: arrest this man, he talks in maths; he buzzes like a fridge, he's like a detuned radio. [radiohead]
As for Moore's Law, well it is more an observation than a law. IMHO "Law" should be reserved for more important stuff Murphy's Law and thermodynamics.
Engineering is the art of compromise.
Code optimization is actually the least of your worries. Most of the latency in modern desktops, for example, comes from memory access, not algorithmic slow-downs.
Try structuring the data better, and you will go far.
I guess algorithm analysis will at some point become more mainstream again. I suppose application profiling will also become more popular.
Interestingly, the available memory will continue to grow, so we might end up structuring our data structures so that access time will be minimal. That is - our data structures will continue to change focus from compactness to raw speed. And big O analysis is part of that picture.
I think we'll see some interesting things happen with fiber technology, though. When those envisioned optimal silicone chips become commonplace and thus really cheap, all appliances might run on them, and thus make it feasible to distribute your processing between your computer, your fridge and your iron. We'll just interconnect everything - perhaps a new fibre connector in our electricity plugs.
Stop the brainwash
If Chip design is at its limit for reduction, then other factors an still come into play. Parallelization and multiprocessing coming to mind. Multiprocessing hasn't reached any type of limit. As chipsets improve, and CPUs play better together, then overall computing power can continue to increase. (Yeah, all you geeks go on and tell me how multiprocessing isn't really doubling and is not as optimized, yadda yadda).
The point is, CPU reduction is not the only path to processing power. It has just been the easiest so far. Watch for other paths to be optimized and utilized as this option peters out.
There is nothing so silly as other peoples traditions, and nothing so sacred as our own.
Honestly, I think a bigger trend will be to take advantage of formalisms that let developers develop more reliable and stable software. Now, I know and you know that things like functional programming have been out there for years, and haven't succeeded because first, they were too slow and therefore wasted too many processor cycles. This is obviously much less of a problem now - Java "wastes" lots of processor cycles, but for a lot of software needs, saves so many human "thinking" cycles that it pays off in spades for businesses that need business or enterprise software to Do Stuff for the back-end sides of industry.
So what big problem(s) are left in the software world? Well, people still bitch about how fucking unreliable most software is. In particular, core, critical system areas, like the interface between hardware and software - as more hardware is out there, and more drivers are developed, and backwards compatibility is an issue, hardware interactions have not become substantially more reliable. And frankly a lot of applications themselves, have become substantially less reliable - the big problem is that adding features and changing GUIs seems to break too many things and introduce too many potential problems (look at Outlook XP vs. Outlook 2000 - fixed some security holes, made a prettier GUI, and made the damn thing crash all the time).
Look at a lot of the academic work being done in computer science, especially in programming language design, operating system design, parallel algorithms and parallel languages. Sometimes researchers head off down dead-end paths, but sometimes they have it right, and it just takes a while for industry to see what they need this stuff for. That being said, it'll always be cheaper to teach people "Programming in Java 101" in India and then hire 1000 of them to hack away at code, admitted usually for the most uninteresting and repetitive types of development work (at least, this will hold until economic parity in the third world becomes a reality).
...what this means in relation to Intel's .09 micron work with Prescott (slated for late next year)? Could be nothing, could be indicative of INtel hitting some stone walls in .09 micron development (which I always knew would be a tough row to hoe for complex cpus.)
Read one post earlier in which the poster thought AMD was abdicating a "clock speed" race. Obviously, this sentiment, among so many like it, comes from Hector Ruiz's speech last week in which he said that AMD wasn't going to do "technology for technology's sake." I wish Hector had made himself a bit clearer...;)
What I think he meant was that unlike Intel with Itanium, AMD was not going to design brand-new technologies with no practical worth simply for the sake of performance (because Itanium has no software it's very nearly useless--except for doing PR benchmarks for Intel.) That's why AMD chose to do x86-64--because it is technology for practicality's sake. That's my take on that statement.
Also, AMD has been out of the "clock race" ever since they designed the K7. The race AMD wants to win, and has been winning, is the "performance race" which doesn't depend on raw MHz. Any P4 will be much slower than any K7, when clocked at the same MHz speed. That's why AMD's been using performance ratings--because they are much better measures of performance than mere MHz speeds could ever be between competing cpus with differing architectures.
Eventually you will reach a limit on the size of the individual swtiches. The one the article gripes about appears to be the sloppy wave function of the electrons letting them tunnel across the junction. But matter is lumpy (quantized) and eventually you'll hit a just-a-few-atoms wall.
... Keep this up until you are done. Laying out your gates for minimum signal run length means you end up with a cube, or something close to it.
But there's more that can be done - in terms of geometry and organization.
Current chips are a single two-dimensional array of components (or sometimes a small number of layers). But build your gates and interconnects in 3-D and you can go farther on two fronts:
- Speeding up the individual functions a bit further. (The more complex, the more improvement).
- Combining a LARGE nubmer of parallel elements into a small space (so they can talk to each other quickly).
Back in the '70s I had a rap I'd do called "preposterous scale integration". Basic idea:
- Use diamond for the semiconducting material (because it conducts heat VERY well).
- Grow a LARGE sheet of it, writing the domain doping and interconnects with ion beams as you go.
- TEST the components as you go:
- Negative power lead is a slow (low accelleration voltage) electron beam.
- Positive power lead is a fast (high accelleration voltage beam) electron beam, causing secondary emission of more electrons than are in the beam.
- Test injection probes are smaller versions of the power leads.
- Test probe is a very slow electron beam, where the electrons turn around at the surface, and a positively-charged region will suck 'em to the chip.
(These are all variants of electron microscope imaging hacks that were in use as far back as the 70s.)
- If a component fails, turn up the current, vaporize it, and deposit it again. Repeat until you have a good one.
- When you're done with the layer, don't stop. Deposit another layer, and another,
- Apply power to two opposite faces of the cube. Use bus bars the size of the cube face - at least near the contact point - to minimize IR drop. Use a good conductor, like copper or silver.
- You need a LOT of cooling. So circulate cooling liquid in the buss bars. (Copper and silver are also good heat conductors, and water is a terrific heat carrier.)
- The other four faces are for I/O. Use semiconductor lasers, photodiodes, and fiber optics light-pipes. You can COVER the faces with fibers. Put your drive electronics and SerDeses in the layer just under the pipes - or dope the index of refraction of the diamond to make a light-pipe into the depths and distribute them throughout the volume.
- Diamond is stable up to very high temperatures, but you need to protect it from air when it gets hot (or it will burn). So put it in a bottle with an inert gas just in case. Limitiing temperature structurally is about where it starts going over into graphite, so you can let it get up to a dull red glow (if your I/O is at some bluer color and that temperature doesn't create too much thermal noise).
- How big can you get? Square-cube law limits your I/O-to-computation ratio, since the I/O is on four faces that go with the square of the linear dimension, the computation goes (approximately) with the volume, or the cube of the dimension. The cooling-to-gate ratio suffers a similar square-cube issue (plus a linear penalty for power losses from the internal distribution busses). You also have an interconnect penalty - as you get bigger you have to give a higher fraction of your volume to power and signal lines (or signal repeaters), but this actually improves the square-cube problems. Finally, construction time is about proportional to number of computational elements. So let's pull a number out of nowhere and say two meters on a side.
Of course the punch line is what the device would look like:
- A six-foot cube of diamond.
- Glowing cherry red.
- In a glass bottle of inert gas.
- Supported by water-cooled silver bus bars.
- And connected to everything else by an enormous number of glass fiber light-pipes.
In other words, the kind of thing you'd expect to be the ship's brain in a late model Sklyark spacecraft, from one of George O. Smith's golden-age science fiction novels. B-)
====
This rap was always entertainment rather than a serious proposal, and is no doubt buggy. For instance: I hear doping diamond is a bit problematic. And these days I'd suggest doing chip-under-construction powering and testing using physical contacts and JTAG fullscan or a variant of the CrossCheck array, rather than (or to suplement) the electron beams.
But I hope the point is made that, for parallizable tasks at least, we still have a LONG way to go with improved geometry before we finally hit the wall.
Bantam Dominique roosters crow a four-note song. Once you've heard it as "Happy BIRTHday" you can't NOT hear it that way