Slashdot Mirror
Intel Devises Chip Speed Breakthrough
Chad Wood writes "According to the New York Times (free reg. req.), Intel has demonstrated a research breakthrough, making silicon chips that can switch light like electricity. The article explains:''This opens up whole new areas for Intel,' said Mario Paniccia, a an Intel physicist, who started the previously secret Intel research program to explore the possibility of using standard semiconductor parts to build optical networks. 'We're trying to siliconize photonics.' The invention demonstrates for the first time, Intel researchers said, that ultrahigh-speed fiberoptic equipment can be produced at personal computer industry prices. As the costs of communicating between computers and chips falls, the barrier to building fundamentally new kinds of computers not limited by physical distance should become a reality, experts say.'"
No req. required
SAN JOSE, California (AP) -- In an advance that could inexpensively speed up corporate data centers and eventually personal computers, researchers used everyday silicon to build a device that converts data into light beams.
Light-based communications has until now largely been the realm of large telecom companies and long-haul fiber-optic networks because of the expense of the exotic materials required to harness photons, the basic building block of light.
Now, researchers at Intel Corp. say their results with silicon promise to reduce the cost of photonics by introducing a well-known substance that's more readily available.
In the study, published in Thursday's journal Nature, the Intel researchers reported encoding 1 billion bits of data per second, 50 times faster than previous silicon experiments. They said they could achieve rates of up to 10 billion bits per second within months.
"This is a significant step toward building optical devices that move data around inside a computer at the speed of light," said Pat Gelsinger, Intel's chief technology officer.
Intel believes the finding could have profound implications for the links between servers in corporate data centers. Eventually, the technology could find its way into personal computers and even consumer electronics.
"It is the kind of breakthrough that ripples across an industry over time, enabling other new devices and applications," Gelsinger said. "It could help make the Internet run faster, build much faster high-performance computers and enable high bandwidth applications like ultra-high-definition displays or vision recognition systems."
Unlike electrons that flow through copper connections common today, the photons in light are not susceptible to data-slowing interference and can travel farther.
The Intel researchers built a device called a modulator, which switches light into patterns that translate into the ones and zeros of the digital world.
A light beam was split into two as it passed through the silicon, which has tiny transistor-like devices that alter light. When the beams are recombined and exit the silicon, the light goes on and off at a frequency of 1 gigahertz, or a billion times a second.
Infrared light is used because it can pass through silicon.
"Just as Superman's X-ray vision allows him to see through walls, if you had infrared vision, you could see through silicon," said Mario Paniccia, a study author and director of Intel's silicon photonics research. "This makes it possible to route light in silicon, and it is the same wavelength typically used for optical communications."
The researchers expect to be able to increase the frequency to 10 gigahertz, making the technology commercially viable, said Victor Krutul, senior manager of Intel's silicon photonics technology strategy.
"This implies that the economies of scale that we have seen for the electronics industry could one day apply to the photonics industry," Graham T. Reed, a professor of optoelectronics at the University of Surrey's Advanced Technology Institute, said in a commentary that accompanied the research paper.
"The device Intel has built is the prototype of a high-speed silicon optical modulator that the company has now pushed above two billion bits per second at a lab near its headquarters in Santa Clara, Calif. The modulator makes it possible to switch off and on a tiny laser beam and direct it into an ultrathin glass fiber. Although the technical report in Nature focuses on the modulator, which is only one component of a networking system, Intel plans on demonstrating a working system transmitting a movie in high-definition television over a five-mile coil of fiberoptic cable next week at its annual Intel Developer Forum in San Francisco."
why do you think there will be size and speed gains?
the complexity of most logical and arithmetic operations that have to be performed on a bit increase exponentially with the number of possible states in the bit.
Disclaimer: I am a Ph.D. in fiber optic physics
This is a 2 Gb/s modulator, whereas III-V semiconductor modulators above 40 Gb/s are commericially available.
A modulator by itself is nothing new, and not the whole story. You need optical waveguides with bending radii much smaller than currently available for routing, and optical logic gates which are an even worse problem.
The article doesn't describe the technology -- is it electroabsorption? Mach-Zehnder?
Nevertheless, a small and fast silicon modulator has obvious commercial value, even if it isn't the greatest thing since sliced bread.
The difficulty with mastering tri-state and quad-state computers (as opposed to bi-state or binary) comes with the gates used. How would one perform an inverse operation when there are two other choices from which to choose? Instead of AND, OR, and NOT (not to mention combinations such as XOR, NOR, NAND, etc.), you would have at least 8 gates (if I recall correctly; I worked on something similar to this during the summer) doing things such as shifting, reversing, "inverting," and such. The different permutations of these make it even more confusing.
In addition to this, you would need to find a medium capable of carrying a tri-state signal (electrons are not best suited for this). In fact, due to the fact that we have a tough time determining on and off sometimes, I would personally suggest we leave it at binary for the time being.
I know it's a long post, but most of it is necessary.
Flourescent and LED lights do generate heat, just not to the same order of magnitude as incadescent lights. Its significantly less, which I specifically mentioned in the post! However there is still some heat generated. If you place a lot of LED lights together though then they can generate enough heat as to become significant.
Fluorescent and LED lights do not get hot.
Sure they do. They are far more efficient than incandescent bulbs, so they produce significantly less heat per lumen, but a very bright fluorescent or LED light can get quite hot.
In fact, high-brightness LEDs like the Luxeon Star have to be mounted on heat sinks to keep them from burning up.
ZFS: because love is never having to say fsck
More info about base 3 computing here.
temperature, is really not the problem. The problem is stabilization. Different gates "stabilize" that is produce consitant output high or low at different rates, gates are strung together into circuits on the chip and thouse circuits then take a certain amount of time to stabilize, this is critical because the output of one circuit will be the input to another be it on the same IC or interfacing with something else. The reason you can overclock is in most cases ICs in computers the CPU in particular are underclocked to begin with. The clock cycle is longer then the stabiliation time when the chip is cool. However the voltage running though the traces and the swiches meets some resistence and part of it is disipated as heat, when silicon-eletric gates heat the respond slower and the stabilization time becomes longer, so the clock cycle must be longer if you want correct output. This is why if you take special meausers to keep the chip cooler you can often run it faster. Fiberoptics are not perfect and can heat too, the smaller you make them that problem is likely to exacerbate. The question I can't answer for you is wether that is a problem at all. silicon-optic gates may not vary in stabilization time in the same way that the electric counter parts do? They may and then the same rules apply or they could have some optimal temp where a cold chip does not work as well as a warm one? It might be they work perfectly up to a certain failure point?
I would love some answers form an engineer who is working with this stuff.
Repeal the 17th Amendment TODAY! Also Please Read http://www.gnu.org/philosophy/right-to-read.html
"Fluorescent and LED lights do not get hot."
This is not true. They do get hot, just not as hot. They don't require as much energy to generate light.
With that said, the question really can only be answered after we know about the design of the chip. If all the light emitting aspects of the chip can be run at full intensity without ever being turned off, and the chip can survive that, then the answer is yes, you can overclock it to the max without it burning out. Will the chips work that way? Well I don't know. We are talking about very small components.
His question was quite valid.
"Derp de derp."
probably the same person that modded yours informative. You are incorrect regarding fluorescents. I can't speak to diodes, but I have known them to be quite hot (such as in a rectifier) so I have doubts about that as well.
Fluorescents DO get hot, as do the ballasts (see post below). I just got done in the lab measuring different ballast systems that use high frequency to energize high output fluorescent lamps. Current generation systems are twice as efficient as older systems by using HF but they still are hot as hell. The ambient temperature of a 100 watt fluorescent lamp, powered by only 65 watts of power (typical cpu power) at high frequency has an ambient temperature of over 100F at 6cm away. The surface temperature is over 212F (100C).
So yes, fluorescents DO get hot. They just produce alot more light per BTU of waste heat, but still hot.
Another problem: fluorescents are plasma devices, similar to neon signs. This means they operate in a semi vacuum (1% of atmosphere), with the electrical fields generated causing an outer electron of the mercury atom to fly off toward the positive end of the lamp, and strike the phosphor coating of the lamp. This reduces the energy in the electron, which then is captured by any mercury atom with an electron missing, thus with a positive charge. This is not a practical solution inside a integrated circuit. This isn't even including the other problems I mentioned in the other post, such as ballasting.
Tequila: It's not just for breakfast anymore!
Can you guys all shut up about Pentium and clockspeed for crying out loud?
This is about optical networking using silicon as the semiconductor. Not about a CPU.
Everyone who doesn't understand what an optical modulator is can go post on the latest SCO story. That is all.
Problem is to have three or four states, you need more complex circuity. Binary is simple and works well. A bit it a gate, a transistor. It's on or it's off, 1 or 0. Well if I want to represent four states, how do I do that? I guess I need to do it by voltage or amperage level. MEans I need a more complicated circut.
Give you something of a parallel in another digital field:
Digital CD audio is stored as 16-bits per sample, 44,100 samples per second. Well that means that to convert the digital data to analogue, which is what sound waves are, you need to change the output voltage of the state 44,100 times per second, and do it to a resolution of 65,536 different levels. Originally, D/A converters tried to do just that, and failed rather miserably. It was just all hell to build a circut that could do a good job of controling voltage that accurately that quick in that fashion.
The answer, it turns out, came from computers and high current variable speed electric motors. Motors of that type are controlled using what is known as pulse wave modulation. Their power source is either all the way on, or all the way off, binary in other words. It pulses at a high rate of speed. What you do is the faster you want the motor to go, the more on pulses you have. Works great, you have a simple design that provides a fine level of speed control. Only down side is the motor whines at the frequency of the pulse.
Now this was applied to audio as well. What you do is convert the PCM data on the CD to a much higher frequency 1-bit PWM stream. That then controls the analogue voltage. It ends up working great, so good in fact that sony has a new system called Sony Direct Stream Digital that just takes and stores the PWM data directly. This type of converter is called a Delta-Sigma D/A converter and is basically the only kind used any more. You may CD consumer equipemnt, espically older stuff (Sony Discmans did it a lot), occasionaly advertise it as "1-bit D/A".
Binary systems are just simpler to implement in electronics, hence we do. It is at higher levels that they start representing data with multiple states.
Heat will probably be a problem. Since you're dealing with photonic crystals, a small change (a few angstroms) in size (heat related) will change the optical properties of the device dramatically. But light doesn't heat up materials quite as dramatically as rapidly switching MOSFETS. And you don't get waste tunneling currents at small sizes either. So you can make better device. However, you CAN'T actually overclock, you'll mess up the optical properties of the device severely if you switch to different frequencies (turning a diffraction pattern that indicates an OR into an AND, for instance).
Karma: Excellent^(-t/Tau), Tau=Wittiness/Trollishness
It is a common misconception that electrons move quickly through conductors. This, however, is not the case. When an electric field is applied to a conductor (e.g. from a battery), the random motion of the electrons in the material gain a small drift velocity. In copper (a relatively good conductor), this drift velocity is on the order of 10^-5 m/s to 10^-4 m/s (much less than c=3E8 m/s). The reason that conductors work the way they do is that the information is carried by the electric field rather than the individual electrons. A good analogy here is to think of a tube filled with ball bearings. Stuff one more bearing in the tube at one end and one pops out of the other "instantaneously". While the inserted bearing didn't travel the distance, it did have an effect at the end of the tube.
Another common error is raised by the parent post. Transmission rate and bandwidth are completely different concepts. The transmission rate refers to the number of bits of information that can be transmitted down a pipe without loss (i.e. the capacity). Bandwidth, on the other hand, is a frequency domain concept and refers instead to the range of frequencies that the pipe can support. While it is true that a system with greater bandwith usually has greater capacity, it is a gross generalization.
Rapidly approaching the Zener knee...
and it seems you have that special sauce investors are looking for down perfectly.
Pah... Save a few bucks and just use the Dilbert mission statement generator
Customize the list of nouns, and you can even make it sound relevant to your own business.
And, for reference, I did actually use that to come up with an "Objective" line for my SO's resume (though as a warning, she works in a field where the resume counted as a formality - she could have used "I want you to pay me to scratch my ass all day" as her objective, and still gotten the job).
If I remember correctly, the optimum base for data size is base e (approx 2.7). I guess that base 3 would be the best we could achieve. Can anybody who knows more about information theory back me up?
Karma: Contrapositive
After reading the article, it turns out that *all* this hoo-ha is about the fact that INtel has worked out how do do telecommunications level optical switching (read LED-LASER-RAPID-BLINKING) on a chip built using "normal" chip fabrication techniques.
This is in no way about "faster CPUs" it's ALL about "now we can fabricate telecomms equipment using standard CPU techniques, so they'll be cheaper and therefore easier to put into devices".
So you're not likely to be getting significantly faster PCs from this technology, though it *does* make more likely the chance of (one day) having a direct gigabit fiber port on your PDA (or digital camera/other-small-electronics-device)
Visit CryptoGnome in his home.
The nature paper
Not even LEDs are 100% efficient. However, for an optical system, the heat production is related to the duty cycle of the lamps, rather than the switching speed, so the heat production should remain constant regardless of clock speed.
That's true of the heat production in the guts of the lamp itself (at a given light intensity). But there are other factors.
On the one hand, this means you don't need to improve cooling to overclock. On the other, it means that you can't improve the overclock level with improved cooling.
Most of the heat loss in a circuit comes from the I-squared-R losses of the currents needed to charge and discharge the stray capacatance of the wiring (even the tiny traces on the ICs) and the space-charge of the devices.
In particular, if the wire has any significant length, you need to run that current through a series resistance (at least at the driving end) matching the impedence of the wire, in order to produce a nice waveshape at the far end and prevent "ringing" as the signal bounces back-and-forth (which would degrade the waveshape at the inputs to far-end gates and make the signal both more sensitive to noise AND more generative of noise to interfere with its neighbors.)
With CMOS you only pull power (except leakage power) when you CHANGE the state of a signal. But when you do, you have to charge, or discharge, the signal wiring through that matched resistance. The impedence of the wiring doesn't change a lot with technology and speed. So with a given length of wire, you have a given amount of energy dropped every time you switch it. Switch it twice as fast, generate twice as many pulses of heat.
New generations of semiconductors fight this in three ways:
- Shrink the components (so they have less stray capacatance to charge and discharge).
- Shorten the signal runs by making the components smaller so they can be closer together (reducing the stray capacatance of the lines). (But this doesn't help for signals that HAVE to cross the chip, or leave it.)
- Lower the power supply voltage (so you don't have to swing it as far. Current goes up with the the voltage, heat loss with the square of the current.) (For signals that leave the chip this may be harder to do than for signals that stay on it - due to external interference.)
For switching a light-emitting device you still have to charge and discharge the capacatance of the device itself and the wiring to it. Switch it faster and IT doesn't heat up much more. But the driver circuit does.
By putting a light modulator on the chip, Intel's new technology wins in two ways:
- You don't have to rapidly switch the power to the laser (which involves switching a LOT of current through an impedence-matching resistor).
- You don't have to run a microwave-speed signal through a long resistive wire, which degrades its waveshape and also produces still more losses.
Instead you switch a low-power, short-range, on-chip wire to a low-capacatance active region on the on-chip modulator. Switching losses are relatively small, comparable to those of a gate-to-gate internal signal in the same chip.
Bantam Dominique roosters crow a four-note song. Once you've heard it as "Happy BIRTHday" you can't NOT hear it that way
In fact the signal on such a wire will tend to hang around at about the level it was last driven for quite a while (the wire is a cap) untill it discharges or some other gate drives it.
In fact internal wires that are genuinely tristate are considered evil in most chip deigns - a floating signal will tend to turn on both the transistors in the gate(s) being driven causing current to flow where it shouldn't (one should be on or the other not both) - chips with internal floating nodes can et into horrible lockupstate which cause thermal runnaway and chip death. Normally if you are using tristate circuits you have a resistor to pull the wire to a known value when not in use, a weak 'keeper' transistor, a protocol which makes sure that someone is always driving them or a combination (PCI is a great example where all the bus clients know whow's driving each wire at any time and when wires are released they are first driven to a safe keeper voltage and then released so a weak resistor can hold them)
My Quake game is limited by physical distance. It takes 100ms to go across the country and back. Latency is the killer here.
Rough, napkin quality calculations here...
m = miles to server = 2000 (round figure for "across the country")
c = miles covered by light in 1 sec
2m/c = 21ms round trip time
100ms - 21ms = time lost to switching hardware, mostly, given that (in my experience) a simple ICMP ping will usually show very similar results, we probably can't attribute it to server processing time.
So, as you can see, there is plenty of room for improvement. Faster/less switching between you and them means less latency. If you have 1/50 second latency, events are reported to you in the time it takes a good CRT to refresh twice.
Light is fast.
Maw! Fire up the karma burner!
Minor nitpick: b doesn't approach Pi, it approaches e. Otherwise, that looks like the formula I've seen in an article on this subject.
Disclaimer: IANAL. This post is, however, legal advice, and creates an attorney-client relationship.
It is fairly uncommon to find transmissions over long distances that are just simple on-off pulses. Even modesms don't do that, and haven't for a long time. They came to find out that 300bps is about the max you can do with simple on-off signaling. So faster modems use more complex modulations that heve multiple different tones and amplitude levels.
On the newest and most abstract level we see DWDM fibre transmissions. This takes multiple signals at different fewquencies of light (the individual transmissions which are usualy more than simple on/off) and multiplexes the singal over a single fibre.
None of that bears any relation to processing on silicon chips.
Perl golf time...
perl -pe 's:(\d{8})\s*:chr oct"0b$1":ge'
Feed it the string of binary on STDIN.
Is there a shorter translator?
what is keeping America afloat?
is a good question
The 8.2% third quarter growth was purchased on credit-the $374 billion budget deficit that was the largest in the country's history. All indications are that next year's deficit will be even larger, exceeding half a trillion dollars.
Any idiot with a hand full of credit cards charged to the next generation's children can gin up the short term illusion of prosperity. Until, that is, the bills come due.
George W. Bush inherited a $127 billion fiscal surplus but ran through all of that and more in his first year. He has turned a $5.6 trillion 10 year forecast surplus into a $3+ trillion forecast loss-an almost unimaginable reversal of $9 trillion in only three years.
The result of this almost psychotic profligacy, according to the Congressional Budget Office, will be a national debt of $14 trillion in 10 years. Interest payments alone will approach a trillion dollars a year and will exceed spending for all discretionary federal programs combined.
http://www.commondreams.org/views04/0105-08.htm
There are places where the networks are not touching,and there are places where they are-Boeing's Lori Gunter