Intel Claims 10Ghz Transistor

Professional Wild-Eyed Visionary writes: "Intel has developed a new CMOS chip technology that cranks out 10Ghz, 400 million transistors per chip, with each transistor only 3 atoms thick, previously thought impossible. See story at Dial Electronics " While this story's rather fluffy, it makes it sound like Intel is a few years ahead of it's earlier projection of reaching 10Ghz by 2005. Of course, maybe they meant integrated into actual chips;) (in which case 2005 still sounds nice).

Comments

Typical Intel strategy

[The Man]

As we've come to expect from Intel, they are doing a tremendous amount of R&D in their fab labs. Intel has some of the best fabrication technology in the business (perhaps only IBM's is better). But when's the last time you heard of Intel actually doing anything revolutionary with that fab technology? That would be 1975. Since then they've only made the same processor over and over again. Seems like such a shame to let all this technology go to waste producing the same processors they made when they launched the company... Will the new IA64 technology be any better? Maybe. Early returns from people working on the architecture seem to indicate that it's just as much of a bitch to work with as x86, but perhaps it's too early to tell. Of course, that technology is 2 years late already anyway.

Re:Typical Intel strategy

[VAXman]

I won't bother showing you the power specifications (which show that P4 uses much less power than Athlon) or the spec and stream benchmarks (which show P4 signficantly outperforming the Athlon, about 2x on specFP, and about 3x on stream).

But as far as the assertion that Intel is an awful place to work, this is easily debunked by looking at Fortune's Best Places to Work list, where Intel is #41, ahead of every computer company including the ones you listed - HP is all the way down at #63, and Compaq and IBM aren't even in the top 100.

Re:Cosmic rays?

[Alien54]

We normally think of cosmic rays as something that causes bit rot (though in practice it's alpha particles). In a chip that has transistors only 3 atoms thick, would this radiation cause physical damage instead? If so, we'd need to think about employing a lossy grid of gates, so that a few failures don't kill the processor

I found this:

"Recently there has been increased emphasis on radiation effects in space due to an increasing number of satellite launches for commercial and defense systems. The natural space environment can damage electronics because of total-ionizing-dose and single-event effects (SEE). These are caused by the high energy electrons, protons, and heavy ions that are intrinsic to the space environment due to cosmic rays and the Earth's radiation belts. SEE due to cosmic rays and high-energy protons can lead to hard or soft errors in many types of devices and ICs. SEE are even possible in avionics and ground applications of advanced microelectronics with submicron feature sizes. SEE can cause failure at any point during a system's lifetime due to one inopportune particle strike, if circuits and systems are not suitably designed, tested, and built. Total dose effects accumulate over a system's lifetime, and can lead to premature performance degradation and system failure."

There are some interesting links on this at the Sandia Labs website here. Some of these go to sites that are a bit encyclopedic.

How much faster can it get?

[dfenstrate]

I was just doing some light math... 1 GHz is 1e9 cycles per second, and the speed of light is 3e8 m/s- so in a single clock cycle of the latest processors, light can only travel 10cm (less, I know, b/c it's electricity, but I'm obviously not an EE) 10 cm is fine for a single chip, but...

When you get up to 10 Ghz, the distance is only 1 cm- and aren't your typical Pentiums and Athlons bigger than that?

So how fast can they realistically improve clock speeds before going back to the drawing board?

Re:How much faster can it get?

[Hater's Leaving, The]

Indeed.
One thing that most people overlook is that Moore's Law is _not_ about processor speed, or throughput, but is actually about _gate density_.
Therefore we should hope to see the functional blocks become smaller as time progresses, so that their output is still available before the next clock edge where it will be routed to the next functional unit in the pipeline.
Note - frequencies have increased faster than densities, so at the moment it looks like it's a losing battle, however, this will simply force chip designers to come up with more fine-grained functional units (and possibly to expect mutiple clock tick latencies between some of the functional units). For example, DEC in their Alpha chips were looking at this kind of design, and AFAIR they were the first people to demonstrate the >1GHz general purpose CPU _many_ years ago (not a production system, a specially cooled unit as proof of principle), which bears out the correlation. (OK, it (the 21164) never reached production at that speed, but what the hell, they had newer chip designs to work on instead).

THL
--

Re:How much faster can it get?

[Soft]

When you get up to 10 Ghz, the distance is only 1 cm- and aren't your typical Pentiums and Athlons bigger than that?

Pipelining takes care of that; even if the information cannot physically cross the chip end-to-end within one clock cycle, what counts is the time required to cross one set of logic gates (one "stage" of the pipeline). Even though one elementary operation takes more than one clock cycle to complete, you can "feed" the pipeline so that you actually get one result every clock cycle except at the beginning.

Re:How much faster can it get?

[Pulzar]

Signals do not bounce around the chip like that -- your "place and route" guys will try to keep things localized, so that register-to-register distances are much smaller than the size of the chip.

Of course, that's not to say that it's not an issue -- it will make things a lot harder to implement. But, then, every time the clock period goes down, it makes things harder. Designers just sit down and find another way to do things faster :).


----------

10GHz Transistors

[commandant]

Actually, academics have created 100GHz transistors out of GaAs. 10GHz isn't that great compared to these ultra-fast ones

However, the distinction may be that this is the fastest corporate-built transistor, and it might be the first semi-integrable one. I don't know the details of either development.

Maybe this is using Si? I forget the frequency limit of silicon, but this may be the fastest silicon transistor ever built.

A new year calls for a new signature.

Re:10GHz Transistors

[OxideBoy]

Actually, a 712-GHz device was published in the early Nineties. It was an antimonide-based nonlinear amplifier (not really usable for switches). 1 THz devices do exist now, but they are based on AlSb-InAs heterostructures and as such are pricey.

Here's an abstract from 1991:

"Oscillations have been obtained at frequencies from 100 to 712 GHz in InAs/AlSb double-barrier resonant-tunneling diodes at room temperature. The measured power density at 360 GHz was 90 W cm-2, which is 50 times that generated by GaAs/AlAs diodes at essentially the same frequency. The oscillation at 712 GHz represents the highest frequency reported to date from a solid-state electronic oscillator at room temperature."

from E. R. Brown et al., Appl. Phys. Lett. 58 2291-2293 (1991).

silicon is relatively slow

[peter303]

It may max out at 10GHz or so.
However gallium arsenide, indium something,
have potential considerably beyond 10GHz and
are being used for high speed D/A and optical
connections. The problem with the non-silicon
stuff is they are harder to fabricate in very
high integration. They tend to be two or more
integration genrations behind CMOS.

Re:atom movement - not true

[JordanH]

• Large relative position uncertainty like you described only applies at the sub-atomic level. An entire atom has a predictable position in space and time.

IANAP (physicist), but I believe that there is some (albeit small) uncertainty with atom positions. I believe that tunneling of hydrogen atoms is how fracto-fusion works. Now, it may well be that it's greatly more probable with a hydrogen atom than a helium atom (and from what I understand, it's not too common with hydrogen atoms), but it does occur.

• Don't worry, your dinner table will never re-materialize a meter from where you were about to set your macaroni.

Never is too strong. There is a finite probability that it could. It might be so unlikely that it would occur, on average, once in 5 billion ages of the Universe, but it could happen.

---

Re:Instantaneous, real-time voice translation?

[djocyko]

Offtopic, but while it is true in context voice recognition is the best approach, it would be possible to incrimentally analyze the sentence as it is being said, thus getting a rough guess at first to having it pretty much set before the last word is said. The last word would solely finish it off. It wouldn't be dead on voice translation, but at the end of the sentence, it would be "instanteneously" recognized.

Re:This is definitely vapour

[Betcour]

The P4 is not a total failure, it's like the Pentium Pro : no software can really show today what the core is capable of. That doesn't mean the core itself is worthless, just that some people need to recompile their apps...

Re:atom movement

[Soft]

Yeah, but the most probable thing always happens, doesn't it?

No, otherwise its probability would be 1. If you prepare a system that has a 10:1 probability to be in a given state (say, you send light and arrange for it to be polarized at about 70 wrt to an analyzing polarizer) and repeat many times the experiment of measuring whether it is in that state (send many photons and detect how many pass through the analyzing polarizer), you'll find it is one time out of ten on average (10% of the photons will get through).

Re:atom movement - not true

[KarmaBlackballed]

Your physics seems rustier than mine.

Large relative position uncertainty like you described only applies at the sub-atomic level. An entire atom has a predictable position in space and time. Need practical proof? Who has not seen the single-atom logo etches IBM and other research departments have been showing over the last decade? Or how about the nano-machines that are just a few atoms thick reported here on /. and other places.

Don't worry, your dinner table will never re-materialize a meter from where you were about to set your macaroni.


~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
~~ the real world is much simpler ~~

Interconnect chips at 50 Ghz

[Nick Arnett]

See http://citeseer.nj.nec.com/294779.html for some research at Stanford on technology that especially applies to building very fast switches. This is standard CMOS at 50 Ghz. The researcher used to work for Intel on the Pentium chips -- I can remember when he was working on trying to break the 1 Ghz barrier... Not sure if the record has been broken, but he told me a few months ago that he had made the fastest processer to date, which was 20+ Ghz. He said built a radar with it...

Pipelining and some random silliness

[Erasmus Darwin]

you can "feed" the pipeline so that you actually get one result every clock cycle except at the beginning.

Mispredicted branches also cause a significant performance drop with pipelining. The CPU doesn't know for sure whether or not it's going to branch until the branch reaches the end of the pipeline. Until then it has to more or less guess based on previously results (or in the simpler case, just always predict "taken" or "not taken") and, if the prediction is determined to be wrong, it must clear out all the partially executed instructions.

Another performance hit is loading data and then attempting to immediatly use that data. Since the load operation takes a couple cycles (memory is relatively slow compared to a CPU), the operation that wants to use that data has to be stalled, creating a gap of a few null cycles between the load and that operation. It's not as bad as a mispredicted branch, but it can be avoided by a smart CPU/compiler combination that places the load operation a few instructions earlier and then works on other stuff while it waits for its results.

Pipelined processors are nifty stuff. It's surprising how conceptually easy a simple one is.

And on a random sidenote, my epiphany on pipelining came when I realized that it's kind of like a fast-food drive-through with multiple windows. A given customer may have a higher latency (because they have to go through that whole start/stop, start/stop non-sense), but the throughput is higher, which sounds like it's only benefiting the store at the cost of the customers. But then I realized that the higher through-put meant that there was less of a backup of people waiting, which benefited the customers. Of course I'm still trying to figure out the corollary for a mispredicted branch. I one day hope to be driving by only to see a little guy in a bulldozer pushing cars out of the line. Then my life will be complete.

Re:Pipelining and some random silliness

[DeeKayWon]

Mispredicted branches also cause a significant performance drop with pipelining. The CPU doesn't know for sure whether or not it's going to branch until the branch reaches the end of the pipeline. Until then it has to more or less guess based on previously results (or in the simpler case, just always predict "taken" or "not taken") and, if the prediction is determined to be wrong, it must clear out all the partially executed instructions.

It is my understanding that IA-64 deals with this by having multiple pipelines so that multiple possible branches can be evaluated simultaneously. That way all that has to be done is to flush the 'incorrect' pipes once the needed data arrives. (This was in an old boot magazine column by Tom Halfhill).

Re:We need smaller atoms

[stevelinton]

There were a series of articles in Nature last summer about the ultimate physical limits of comnputation. Specifically one of them looked at the ultimate 1kg laptop. If you want serial processing, then the answer is a very carefully structured 1kg black hole, which can do about 10^16 quantum bit operations on a 10^16 qbit "word" in the 10^-19 second lifetime of the black hole. Of course the power consumption, cooling and containment problems are rather severe. If you don;t play clever games with reversible computation and very very strong mirrors, then you need a supernova to power the thing.

Cosmic rays?

[FTL]

We normally think of cosmic rays as something that causes bit rot (though in practice it's alpha particles). In a chip that has transistors only 3 atoms thick, would this radiation cause physical damage instead?

If so, we'd need to think about employing a lossy grid of gates, so that a few failures don't kill the processor.
--

Re:10Ghz transister, not CPU!

[keesh]

IIRC, there have been 8GHz transistors (or mosfets) available for a few years now. Nowhere near that small, but they exist. I think this is more a publicity stunt from Intel, trying to claw back some custom from AMD.

Dupe?

[phaze3000]

Is it me or does this sound very similar to the article Intel Says 10GHz By 2005?
2005 is just too far away for me to get excited anyway..

--

atom movement

[roguerez]

My physics is a bit rusty, but if I'm not mistaken these 3-atomic layer thick transistors must have some problems because at this level the predictability of atom movement comes into play.

Every atom has a certain frequent movement. Objects consisting of a large number of atoms stay in one place because the movement of all those atoms combined adds up to zero.

Theoretically, it's not impossible that your dinnertable would suddenly be a couple of meters away from its original place. But it's the statistics that make such an event impossible in practice.

When creating objects very small - consisting of only a few atoms - the movement of every atom get's more important. Chances that the movement of one or more atoms influences the behavior of the object itself (in a way that its behavior is not predictable anymore) are a reality when creating transitors this small.

Therefore I'm amazed by the comment of the Intel scientist that these transistors behave just like other - bigger - devices.

Re:atom movement

[Soft]

But at a subatomic level, it's already dettermined which way the light will be polarised. It's like tossing a coin, you think it's random but at the atomic level it's predetermined which way it goes.

Not according to quantum physics, which states that particles are probabilistic even down to the subatomic level. What is deterministic is a system's wave function, which yields the probability of the system being in a given state at a given time.

Maybe you are referring to the "hidden variables" interpretation, which is quite controversial and almost debunked (see this "Layman's guide to quantum physics").

Re:atom movement

[Fnkmaster]

Wow. No, not at all. Quantum Mechanics is the set of physical sciences based entirely on the supposition that in fact atoms and subatomic particles actually behave probabilistically. Or rather, a series of observations in the early 20th century led physicists to no other reasonable explanation of what was happening in a wide variety of experiments - accepting into their theory base that in fact atomic events are basically probabilistic allowed the derivation of a wide variety of phenomena.

Macroscopic samples may contain a large set of items, depending on size and type of measurement may generally be on the quantum (probabilistic) or macroscopic (observably deterministic) scales. But the macroscopic statistics don't affect the fact that when you get down to the quantum level, to the best of modern science's ability to explain, things are not deterministic.

For the sake of edification, there are theories called hidden variable theories in quantum mechanics that attempt to remove probability from microscopic systems and posit that in fact we simply have insufficient knowledge about the way such systems really work. No such theories have been adequately proved to this point in time.

Re:atom movement

[fatphil]

Sorry, you are utterly off base.

This is why Quantum Mechanics caused such a stir when it was first posited. Even some of the best minds in the world refused to believe that the state of something could remain undecided.
The spin/polarisation/whatever _is_ unknown, and is described by a complex (x+iy) probability function, only upon measurement does the spin/whatever briefly enter a known state, but this precision starts to fade instantly. The real probability of it being in a particular state is the absolute value, or amplitude, |x+iy| = sqrt(x^2+y^2) of the complex wave function.

For example, it has been shown that you can artificially keep particles with a constant spin by continually testing their spin. As you test it you get a true/false result, meaning that you've either got the spin you want, or you have the opposite. If you test it again almost instantly, the wave function hasn't had enough time to make the opposite state particularly likely, and so you almost always get the same spin result, time after time after time.

FatPhil
--

Re:atom movement

[esonik]

The chances of movement of an atome reduce drastically with decreasing temperature. There is an energy barrier that has to be overcome in order to make the diffusion step. For Oxygen in Silicon this is approx. 2eV which is very high (need several hundred degrees Celsius to trigger diffusion). In fact, one way to make thin oxidized Silicon films is to expose the Silicon surface to Oxygen and heat it (several hundred deg.). The oxygen will diffuse into the Silicon and form the oxide. The layer thickness depends on temperature and duration of the treatment. However the article does not say whether they used this technique to get the gate oxide.

Re:atom movement

[fatphil]

Einstein Poldovsky Rosen Hidden Variable Theory leads to contradictions which can be physically demonstrated to disprove the theory. The "EPR" experiment, it is called, named after those three.

FatPhil

--

10Ghz transister, not CPU!

[stripes]

A 10Ghz transister can only make a 10Ghz CPU if each pipeline stage (plus sync overhead) is only a single transister. Which is pretty impossable (a simple flip flop is several transistors, an adder is a big pile of them). As I recall the failed 500Mhz PowerPC that some compony like "eXponential" was making was thought to be extreamly aggressave with only 50 or so transitor delays between pipe stages (and some pipe stages were mostly wire delay to get the signals from one part of the chip to another!). Or maybe I'm confusing that with sombody or others barrel processer style MediaCPU (also out of bisness).

Tiny transistors are wonderflu. Tiny fast transistors are more wonderful. But 10Ghz transistors are no where close to letting you make a 10Ghz CPU. In fact it might be slower then current state of the art (but smaller). Something in this story doesn't add up.

Re:10Ghz transister, not CPU!

[df1m]

They don't say the transistor runs at 10GHz, they say it is very small, and will allow the creation of chips that run at 10GHz.

- dave f.

Re:10Ghz transister, not CPU!

[Velox_SwiftFox]

But the actual article referred to doesn't say "10 GHz transistor", merely that it would open the door to building 10GHz CPUs.

So the transistor might be much faster.

Re:10Ghz transister, not CPU!

[Pulzar]

You're thinking of pipelines, but there are no pipelines with 1 transistor wide stages :).

Which doesn't really matter, since the article is talking about enabling CPUs to work at 10GHz, not about 10GHz transistors.


----------

Re:10Ghz transister, not CPU!

[billcopc]

Why doesn't someone just take a P4-2ghz and remark it as 10ghz ? That's basically what Intel's going to do since they have lost the true definition of performance anyways. Just look at MMX, MultiMedia eXtensions; more like Masturbatory Marketing eXperiment. They will come out with another gimmick that runs at 40 bazillion hertz yet takes 500 clocks to do an XOR thanks to wait states.

At this point, instead of increasing cpu speed, I feel it would be better to focus on SMP and tweaking the other parts of a PC. A faster northbridge and better system bus would relieve important bottlenecks in the PC architecture with a much greater impact than processor speed alone. Just look at all the speed freaks (myself included) who prefer to keep an older cpu but upgrade everything else.

Here's my personal example : I have a Celeron 566 @ 850. Slow according to today's standards, but good enough for everything I do and it runs every game out there. On the other hand, I've got a Geforce2 GTS, 512mb ram, IDE-Raid 80gb, 8x Plextor cdr, Boomslang mouse, yadda yadda. My main CPU is just decent but I cranked everything else to the max. Instead of having a fast cpu that just spends more time waiting for the hardware, I have fast hardware that lets me work and play faster without needing to stay on the bleeding edge of AMD/Intel wars. Sure, I could visit my little chinese dude and his grotty parts shop to pick up an Athlon-C 1.2ghz with mobo and pc266 ram, and my Q3 framerate might jump from 110 to 130, but overall will I get anything done faster ? Only marginally, since the cpu will spend more time snoring while every other component chugs along as the same speeds since 1994. On the other hand I'll be the first guy in town to get my hands on a Geforce3, not because I'm a gamer, but because it will make a more noticeable _perceptive_ difference than spending the same amount on a CPU. I don't care what the benchmarks say, my PC feels just as slow as my P200 did three years ago, and my 486 before that. It's everything else that's been steadily progressing over the years; video, disk, memory, sound. That's what really makes a difference to my eyes and ears. GHz-wars just look good on paper and in Intel's bank account, nowhere else.

Re:10Ghz transister, not CPU!

[smartfart]

I think they meant 400 million transistors running at 10GHz each, which would add up to a CPU running at 4 billion GHz. Of course, real geeks will be able to overclock this baby and have it running at 4.5 billion GHz or more.

This is definitely vapour

[whanau]

Intel realise that they are no longer the kings of the chip game. With their recent P4 release being a total failure, it is only a matter of time before AMD takes over their current position in the market. Releasing this kind of "news" only shows that they are simply trying to play the pr game, rather than actually focusing on proper R and D like AMD and Transmeta

Interference at 10GHz

[Mr_Dyqik]

This is the frequency band that mobile phones use (GSM 900) so couldn't there be problems with interference, and public hype along the lines of mobile phone radiation.

Also at these sort of frequencies you have to use microstrip waveguides to carry your signals, as standard wires don't work so good, so would interconnects and the like have to be redesigned?

Anyway, most computers are limited by memory bandwidth nowadays, and 10GHz chips only makes this worse. To get performance up a lot it would probably be better to improve the memory clock by a factor of ten than the raw processor speed.

Re:Ten gigahertz?

[generic-man]

I trust independent research labs like Advanced Prototype Packet-Layer Engineering to do my benchmarks. They do quality work.

Re:Interference at 10GHz should be slight.

[Jeff DeMaagd]

IIRC the speed of electrons in copper is about 0.3c. But I have forgotten the source of that data though, so if you are trying to make useful calculations that data should be looked up.

some thoughts

[esonik]

Making flat structures (gates oxides) 3 layers thick isn't that hard. What's hard is to make them that thick over the whole wafer and to make a working transistor (they claim the latter). The lateral structures are 30nm which is approx. 100 atom layers wide. Reducing lateral structure size is a lot harder.

One transistor is not many

[Soft]

They may have done one transistor (and actually I seem to remember that IBM had succeeded with a single-atom one), but for doing anything useful you have to pack several of them together... And the closer you squeeze them and the faster you ask the electrons to get between them, the more you are subject to the tunnel effect, that is, the less said electrons care about the paths you carefully etch for them.

Indeed, the more energy they have and the thinner the isolation between "wires", the easier it gets for them to "hop" over the latter. By then anything can happen, bits leaking from one memory cell to the next, calculation errors...

They may be on the right path, but the way to go is quite long.

Instantaneous, real-time voice translation?

[TeknoHog]

Smaller transistors are faster, with Intel claiming the device could eventually pave the way for science fiction technology such as instantaneous, real-time voice translation.

Somehow I can't see how the speed of the transistors can help with the fact that you usually have to wait until the end of sentences before translating, you cannot just do it word by word.

--

Re:We need smaller atoms

[joto]

How about increasing gravity? A black hole computer? I guess we would need wireless networking for this to work...

AMD

[Perdo]

Unless AMD gets to 10Ghz in 2004 in which case Intel will release a 10Ghz chip the next day, availability limited to 10 chips, half of which Intel will keep for developmental purposes.

Re:atom movement - not true

[zaius]

You're entirely correct, Heisenberg's Uncertainty Principle states that we can only come up with probabilities for the positions of sub-atomic particles (electrons, protons, neutrons etc.). AFAIK, there is nothing to suggest that atoms do not have fixed positions.

At most, your table might rearrange itself enough to let a couple protons through, although even that is highly unlikely.