Slashdot Mirror
Intel Researchers See Moore's Law Becoming Obsolete
prostoalex writes "A paper, published by Intel researchers, claims we might be the witnesses of Moore's Law becoming obsolete, as the rate of shrinkage for transistors goes lower with each year. In 2018 we might be able to get the chips manufactured with 16-nanometer technology, then one or two more manufacturing processes will shrink it even further, but after that we're facing the physical limits."
Silicon is, indeed, close to its limit, but that does not mean semiconductors are.
This Wired article, which I'm sure many of you have read, details how new industrially-produced diamonds, thanks to their cheap price and purity (most importantly, being absolutely identical to each other), along with research done by both the government, several corporations, and possibly Intel, may make unbelievably fast systems powered by diamond semiconductors possible.
Some interesting quotes:
Also, a rather ironic one from Intel themselves:
Silicon is dead. Long live diamonds!
looks like they're gotting slashdotted like Kathleen Fent on her wedding night...
Dec. 1 -- Moore's Law, as chip manufacturers generally refer to it today, is coming to an end, according to a recent research paper.
GRANTED, THAT END likely won't come for about two decades, but Intel researchers have recently published a paper theorizing that chipmakers will hit a wall when it comes to shrinking the size of transistors, one of the chief methods for making chips that are smaller, more powerful and cheaper than their predecessors.
Manufacturers will be able to produce chips on the 16-nanometer manufacturing process, expected by conservative estimates to arrive in 2018, and maybe one or two manufacturing processes after that, but that's it.
"This looks like a fundamental limit," said Paolo Gargini, director of technology strategy at Intel and an Intel fellow. The paper, titled "Limits to Binary Logic Switch Scaling -- A Gedanken Model," was written by four authors and was published in the Proceedings of the IEEE (Institute of Electrical and Electronics Engineers) in November.
Although it's not unusual for researchers to theorize about the end of transistor scaling, it's an unusual statement for researchers from Intel, and it underscores the difficulties chip designers currently face. The size, energy consumption and performance requirements of today's computers are forcing semiconductor makers to completely rethink how they design their products and are prompting many to pool design with research and development.
Resolving these issues is a major goal for the entire industry. Under Moore's Law, chipmakers can double the number of transistors on a given chip every two years, an exponential growth pattern that has allowed computers to get both cheaper and more powerful at the same time.
Mostly, the trick has been accomplished through shrinking transistors. With shrinkage tapped out, manufacturers will have to find other methods to keep the cycle going.
These issues will likely be widely discussed this week, when the International Technology Roadmap for Semiconductors is unveiled in Taiwan. The ITRS, which is comprised of several organizations, including the Semiconductor Industry Association, outlines the challenges and rough timetable for the industry for 15 years. A new version of the plan will be released in Taiwan on Dec. 2.
Still, Gargini said, researchers are exploring a variety of ideas, such as more efficient use of electrons or simply making bigger chips, to surpass any looming barriers. Other researchers likely will dispute these conclusions.
"We cannot let physics beat us," he said, laughing.
THE DISTINGUISHED CIRCUIT
The problem chipmakers face comes down to distinction and control. Transistors are essentially microscopic on/off switches that consist of a source (where electrons come from), a drain (where they go) and a gate that controls the flow of electrons through a channel that connects the source and the drain.
IT Jobs
Free It Downloads
Learn Software
Harry Potter Gifts
When current flows from the source to the drain, a computer reads this as a "1." When current is not flowing, the transistor is read as a "0." Millions of these actions together produce the data inside PCs. Strict control of the gate and channel region, therefore, are necessary to produce reliable results.
When the length of the gate gets below 5 nanometers, however, tunneling will begin to occur. Electrons will simply pass through the channel on their own, because the source and the drain will be extremely close. (A nanometer is a billionth of a meter.)
Gargini likens the phenomenon to a waterfall in the middle of a trail. If a person can't see through it, they will take a detour around it. If it is only a thin veil of mist, people will push through.
"Where you have a barrier, the electrons penetrate a certain distance," he said. "Once
Perhaps Intel could hold off on the 10 ghz chips and concentrate on making some that don't get so damn hot.
Oh, Edmund, can it be true? that I hold here, in my mortal hand, a nugget of purest green?
This article has some interesting "facts" about how transistors work. I particularly like the following quote:
This is amazing. MSNBC has apparently re-written everything known about current, and logic sensing! As any undergraduate Electrical Engineer could tell you (and quite a few other people too), current flows against the direction of electron flow, not with it. If electrons are going one way, current is going the other way. That's been the convention for a VERY long time. Current is positive flow, not negative.
The other somewhat amazing claim here is that there is a logic "1" when the transistor is on and allowing electrons to flow, and a logic "0" when it is blocking them. That's amazing to me, since actually, it's the voltage at any given spot that determines the logic, not the on/off state of the transistors. And actually, one of the main benefits of CMOS technology is that between clock cycles when nothing is happening with the circuit (it is static), it consumes almost no power since no current is flowing. Charges exist, and some transistors are "on" and others are "off", but no current is flowing! (Note to other EEs: Yes I know that at current blindingly fast clock speeds, this benefit is largely gone, since few logical cells at any given time are actually not switching and charging up/down, but that was the original idea.)
Oh ya. The last thing is that in NMOS transistors, the electrons do flow from the Source to the Drain as the article said, but in PMOS, they flow from the drain to the source. And it's the Gate-to-Source voltage that's important, not just a voltage applied to the gate.
I wish they had somebody with any engineering skill, or at least a basic understanding, or at least run the article past somebody with some basic understanding of this. The writer of the article obviously has no actual knowledge whatsoever.
Erioll
4th Year Undergraduate Electrical Engineer
20nm marks the edge of the soft X-ray band in the energy spectrum and thats not a good thing to put into people's homes. Those freqencies would make working with your case open very dangerous and proper shielding would become pretty important. It's bad enough we're regularily dosed by low level X-ray emissions from CRTs but once we hit that 20nm range we're talking about harmful radiation exposure.
Also the weight of laptops would increase dramatically once lead shielding becomes a requirement...
"Quando Omni Flunkus Moritati" -- Red Green
Moores law was origionaly "the number of transistors on a given amount of integrates circut space will double every 12 months". It has been basterdized twice, first changing density to speed, and secondly changing the timeframe from 12 to 18 months.
It is truly astounding how much bull shit is heard in this place. Please don't cloud the landscape with any further utterings of your ignorance.First of all, no one ever said that a phone line couldn't handle more than 2400bps that knew what the hell they were talking about. There was a limit in hardware at the time, but the bandwidth of the line was never in doubt.
64k is the very real limit for information transmission within the voice-band over standard copper pair used for POTS (plain old telephone service). 56k is the practical limit with 8k being used as overhead. 64k total is possible on ISDN because it uses out-of-band signalling (i.e. nB+D systems where n is the number of data channels, and D is the signalling/control channel) and thus doesn't violate the 64k limit. Any further realization of higher bit rates is either due to compression or transmitted in a different frequency range (i.e. xDSL)
From Intel's website: "Moore observed an exponential growth in the number of transistors per integrated circuit and predicted that this trend would continue. "
To be a tiny bit pedantic, Moore's original paper talked about the number of transistors per integrated circuit at any given price point. You can always stick more transistors on the chip if you're willing to throw sufficient amounts of money at the problem, but to get those transistors for a reasonable price is another matter.
FWIW, Moore's original hypothesis was that the transistors/$ would double every 12 months, so his "law" hasn't been correct for quite some time. We had been seeing a doubling of transistors about every 18 months for a while, but now it's more like every 24 months. With the current troubles that Intel, AMD and IBM all seem to be having at implementing their new 90nm manufacturing process, it seems likely that the pace will continue to slow.
Huh? I think you got that backwards -- smaller gates require lower voltages (allow, really, since we like it when we can use lower voltages -- it saves power and makes switching faster.)
:)
If you think about it a little, old (big) chips were 5V (remember that?), then 3.3V hit around the PCI era (in those days, I/O voltage and internal voltage we usually the same.) Then 2.5V (often with 3.3V on the I/O still), and 1.8V, etc. As the process geometries have shrunk, they have used lower and lower voltages.
If you still don't believe me, try applying a significantly higher voltage to one of your CPUs. That makes the transistors run better, right?
everything in moderation
I see that Intel finally got around to reading The Age of Spiritual Machines by Ray KurzweilChapter 1, (published in 2000, I might add)
;)
"So Where Does That Leave Moore's Law?
Well, it still leaves it dead by the year 2020. Moore's Law came along in 1958 just when it was needed and will have done its sixty years of service by 2018, a rather long period of time for a paradigm nowadays. Unlike Moore's Law, however, the Law of Accelerating Returns is not a temporary methodology. It is a basic attribute of the nature of time and chaos -- a sublaw of the Law of Time and Chaos -- and describes a wide range of apparently divergent phenomena and trends. In accordance with the Law of Accelerating Returns, another computational technology will pick up where Moore's Law will have left off, without missing a beat"
Down to the exact date! Well, at least they caught on before it was too late
but we are already switching the fuel technology backbone to Hydrogen
Hehe, I always get a kick out of it when people start talking about our new "hydrogen based society" or some other garbage like that. It's incredible how many people seem to believe that you can generate power from hydrogen! Of course, anyone with an once of scientific knowledge can tell you, unless you're talking about nuclear fussion, than hydrogen is simply an energy carrier and not an energy source. You don't pick hydrogen off the magic hydrogen tree, you don't mine hydrogen from the ground and it definitely doesn't just materialize. You put energy into water, you get hydrogen and oxygen. You combine the two back together again at a later date and you get most (but not all) of the energy back. Long story short, you've basically made a cell (aka a "battery" in commonspeak). It's no coincedence that we call these things "fuel cells".
There may be ways to break down hydrocarbons cleanly, efficiently and *cheaply*, thus providing another source of hydrogen where you can get more energy out than you have to put in, but guess where those hydrocarbons come from? If you said, oil, you win the prize!
In any case, in a vain attempt to bring this back on-topic, nanotubes and the like do provide some interesting new long-term possibilities for producing ICs, but they are definitely not without their own set of constraints. No matter how you slice it, sooner or later you run into a minimum size. At some point in time you just don't have enough atoms left to keep your electrons where you would expect them to be. There's lots that can be done in new and different ways to help push these problems further back, but no matter what technology you chose you eventually hit the same sorts of limitations.
Long story short, don't hold your breath for nanotechnology to revolutionize ICs, and definitely don't hold your breath for a society "powered by hydrogen"!
Interesting, but there are a lot of other issues at work here. Take, for example, memory bandwidth.
100MHz Pentium had ~ 533MB/s of memory bandwidth
3.0C P4 has 6400MB/s of memory bandwidth
533MB/s / 100MHz = 5.33B
6400MB/s / 3000MHz = 2.13B
As you can see, memory bandwidth has only increased half as quickly as your processor speed and memory size (actually it's not quite that bad since the P4 reaches a higher percentage of it's theoretical peak than the old Pentium does). But it gets worse.
100MHz Pentium had ~ 300ns memory latency (rough guess here, I can't find any exact numbers, but it's in this range). 3.0GHz P4 has about 75ns memory latency.
300ns / 100MHz = 0.003 s^2
75ns / 3000MHz = 0.000025 s^2
Now THAT is a real killer, and the main reason why things like cache and memory prefetching have become such a big deal. Heck, even cache latency has become a big deal since you could easily wait for as many clock cycles to get data from cache as you used to wait to get data from memory. At 100MHz, you clock cycles are 10ns long, so you only need to wait for 30 clock cycles to get data from memory. For comparison, the L3 cache of the P4EE/Xeon has a latency of about 30-40 clock cycles.
You also get some similar numbers if you look at hard drive bandwidth and latency. Our hard drives our quite a bit faster now than they used to be, but as a fraction of the processor clock speed they are MUCH slower, particularly when you're talking about latency (roughly equal to seek time in hard drive speak).
Or at least it should be obvious. Claiming that "Moore's law is not obsolete now" != "Moore's law will go on forever".
Sean