Slashdot Mirror
Intel Announces Lasers On a Chip
wonkavader writes, "The New York Times reports that 'Researchers plan to announce on Monday that they have created a silicon-based chip that can produce laser beams. The advance will make it possible to use laser light rather than wires to send data between chips, removing the most significant bottleneck in computer design.' The work is from Intel and the University of California, Santa Barbara. This suggests breakthroughs in both computing performance and networking." From the article: "The breakthrough was achieved by bonding a layer of light-emitting indium phosphide onto the surface of a standard silicon chip etched with special channels that act as light-wave guides. The resulting sandwich has the potential to create on a computer chip hundreds and possibly thousands of tiny, bright lasers that can be switched on and off billions of times a second." Further details in the Intel press release.
We're already halfway there. How long can it be before someone makes the frikkin' obvious next development?
nothing, instead of EM pulses propigated by electrically conductive substances, it will be self propigating photons directed by optics.
If I'm reading it right, most of the control could be handled by the same mechanisms, it's just that different signal senders and recievers will need to be used.
And, I thought lasers didn't offer significantly lower latency, only better bandwidth?
34486853790
Connection too slow for X forwarding? Try "ssh -CX user@host"
This looks more for CPU interconnects than for actual CPU processing.
The data still has to be transmitted and still has to get back.
liqbase
Electrons in a wire move very slowly. Inches per hour. Current travels at c.
When you are talking about electrons you start to have problems with resonance and interference between the connections. This is why memory is such a difficult problem, because manufaturer A has to create a memory module that plays nice on the generic memory bus designed by manufacturer B. If there is an optical buss from the CPU to the memory module, the memory manufacture has carte blanch to design a module as fast as they want, because there is no more buss restrictions. They would only have to solve the electrical interference on their module, and hopefully would eventually go all optical.
Please sign petition to restore sanity to our banking system!!!
http://financialpetition.org/
Electrons do travel slow. I don't know if its 6 meters per second, but that's the right order of magnitude.
But the signal is still transmitted by the electrons, not some EM pulse. Most designers try to minimize the EM radiation. Think of it like a tube full of marbles. If you shove a marble in one end, one will immediately pop out the other end... it doesn't matter that it would take a long time for that specific marble to travel to the other side.
I'm in the industry, but this isn't my specialty. From what I remember, the speed of the electrons isn't why this is important. There are electromagnetic effects that limit the speed of communications... things like crosstalk. The little balls, wires, or deposited metal that they currently use to make the interconnections are like tiny little antennas. The interconnections are also a pain in general, no matter what technology is used, because of things like thermal mismatches and encapsulation problems. From a packaging standpoint, this would solve many problems, and probably create even more - alignment, anyone?
W..w..W - Willy Waterloo washes Warren Wiggins who is washing Waldo Woo.
I was at a conference last weel (http://www.ieee.org/organizations/society/leos/LE OSCONF/GFP2006/index.html) were this was presented by John Bowers. As they explain briefly in the article, they are bonding InP to Silicon wafers. The silicon provides the waveguiding, and enough of the mode is in the InP to give them gain. They achieved an optically pumped laser, and were still working on an electrically pumped one. I wonder if this announcement will mean that they achieved electrically pumped lasing.
It's good work, but I'm not sure if the bonding process will ever be suitable for monolithography integrated CMOS and photonics. I was more impressed by the work done in Huffaker's lab (http://www.chtm.unm.edu/huffaker/index.html) where they are working on growing III-V materials directly on silicon. However, the work by Bowers is more mature and will lead to devices sooner.
Uh, Apple uses Intel. Heard? Also, you don't need floppies for XP. Except for some corner case--you can boot CD's or thumb drives just fine.
People are concerned about bandwidth, not speed. I.e. how much data can you put down a wire (and how big is the wire). Or, at least bandwidth is the only thing they can hope to improve, electrical signals already travel pretty close to the speed of light. Part of what limits electrically lines is RC limits - frequencies beyond resistance * capacitance can't travel. Any line is going to have finite capacitance and resistance. In addition, there may be dispersion and other effects causing high frequency pulses not to travel well.
Because the frequency of light is so high (~200 Terahertz at a wavelength of 1550 nm), light can carry a lot more bandwidth before similar troubles set in. But, making transistors and wires is easy. Making lasers, modulators, detectors, waveguides isn't.
What this does is make it much simpiler (and CHEAPER) to make the laser light, to the point where it's worth while to have a fiberoptic connection between, say, your CPU and and your vRAM, or between your IDE controller and your RAM, rather than the terribly capacitive and inductive (and therefore SLOW) motherboard trace.
Those who fail to understand communication protocols, are doomed to repeat them over port 80.
The signal is not transmitted by electrons--if it was like a tube of marbles, it would take minutes to turn on a light switch and seconds to get a single byte off of an external hard drive, which is obviously not the case. The signal is transmitted by voltage differences, which do change and propagate at a rate very close to c.
-- listen to interesting music, support independent radio... WPRB
No. Not like that. That uses compound semiconductors like GaAs (gallium arsenide).
Intel is now making lasers with silicon substrate.
However, if your point is that is isn't quite new, OK. Intel announced this originally back in February 2005 [http://en.wikipedia.org/wiki/Raman_laser]
...the future crusty old bastards are already drinking the Kool-Aid.
Good points you have there, and for further probing, here is an excellent article on the topic from the always excellent IEEE Spectrum:
The Silicon SolutionThe view was horrible and the smell was even worse; Julie severely regretted becoming a proctologist.
Electronic signals travel pretty damn close to c. The problem is that electrons are fermions and as a result are antisocial by the Pauli exclusion principle no more than 2 in each location. Charge makes this even worse. On the other hand photons are boson and they like to hang out in the same location. As a result electrons are handy when you want bits to interact (logic gates, memory) while photons are handy when you want bits to pass through each other (communications etc.). The advantage of using photons is that you can make connections without EMI or other cross talk problems. In addition there is some very nifty quantum computing you can do with such systems (the topic of my dissertation).
One huge advantage could be an orders-of-magnitude reduction in the current necessary to drive signals off-chip. (It's not mentioned in the article whether these drivers have a power advantage) Off-chip drivers are a significant source of current drain in a chip, and if this technique eliminates the necessity to wiggle the off-chip capacitive loads at high frequencies, then you'll see much lower power. And if each pin on the output bus is drawing less power, you may see larger bus sizes and more bandwidth between chips.
Who do you get to be an expert to tell you something's not obvious? The least insightful person you can find? -J Roberts
The speed of electrical propagation in copper (~200,000 km/sec) is about 2/3rds the speed of light in a vacuum (~299,792 km/sec). Think of it as having about 2/3rds the latency of copper and you'll be about right, assuming the light goes through open air.
Now if you mean light through an optical cable, it's about as slow as a signal through copper, so there's no real gain.
The real benefit here is short interconnects without any medium in-between. CPU vendors have done this within chips by putting edge contacts on cores so that they can tessellate the cores and have them connected together. With optical edge connects, the failure rate will be lower because the contacts won't corrode and don't have to be soldered.
Check out my sci-fi/humor trilogy at PatriotsBooks.
Actually, you are both right and wrong.
The old anouncement uses Raman gain- where you throw shitloads of optical power down a waveguide (or fiber) at one wavelength and you get new light at another. For this to happen, you only need a silicon waveguide (and perhaps some electronics to pull out carriers that are formed).
In this new case they are bonding InP (a III-V material like GaAs) to silicon. This hybrid device allows the light to be guided mostly by the silicon, but the gain is occuring in the InP in a typical way.
You are correct, the average velocity of a given electron in a DC circuit is pitifully slow. I think it takes an hour for an electron to make it from the battery through the starter switch and into the solenoid. This is because the electron starts to take off, runs into an atom and bounces backwards like a bouncy ball, hits something else and bounces forward, etc. Hence why we discuss the average velocity. You might also want to look up drift velocity.
However, the electromotive force (emf, colloquially referred to as voltage) propagates as an electromagnetic wave. The speed that it propagates at is dependent on the permittivity of the material it is propagating through.
IIRC from my VLSI class, if you take into account the permittivity of silicon, electrical signals (emf; voltage) propagate at approximately 2/3rds of the speed of light.
:(){
However, a changing voltage signal will propagate at speeds of order c (smaller than c, of course). The 'wires' ror traces unning on the microprocessor are basically transmission lines, so you're really transmitting electromagnetic signals. This is just like standard textbook transmission lines (eg, coaxial cable or waveguide). And of course in those cases, even at low frequencies of MHz waveforms, you're really sending photons, which are nothing but quanta of electromagnetic radiation, down the transmission line. Very long-wavelength photons, but still photons none-the-less.
The limit of the speed of light (or of signal propagation) is one reason CPU's need to be small, so various transistors can talk to each other within an appropriate number of clock cycles. Another very important reason is that every little trace, or wire, on the CPU itself is a transmission line, and as such has its own self-inductance. It also has mutual inductances between other lines, as well as capacitive coupling between ground planes and other devices. Thse parasitic capacitances and inductances act as low-pass filters, effectively reducing the bandwidth of the transmission lines. So chip designers, in the push for more GHz, are always trying to reduce these parasitic elements by making their devices smaller and smaller.
What this latest research probably implies isn't necessarily much in terms of a single electronic CPU going much faster, but with future advances in optical signal processing, it can allow optical elements to be grouped closer together and allow for faster optical processing. Additionally, it may increase the bandwidth for signals from CPU to optical transmission lines (eg, fiber optics) by grouping them more tightly to the processor.
make world, not war
Yes and no, the signal is actually photonic in nature, it's an electromagnetic oscillation travelling down the wire, which itself is nothing more than a simple waveguide. So you're sending photons down the wire, photons being the 'particles' exchanged by two electrons that exhibit Coulomb repulsion.
make world, not war
The simplest way to explain this is to note that a wire is an inductor - and at high frequencies this matters. What is more, a 1Ghz digital signal needs bandwidth much larger than 1Ghz - or the edges of ones and zeros get distorted too much. If CPUs used analog signals inside to transmit information between chips (like a miniature wireless card) one would get similar speed, but this is hard and requires antennas much larger than a single transistor.
With light one gets the best of both worlds - the laser beam is analog, coherent medium which is modulated with a digital signal. So you can use a waveguide to distribute it, but a "simple" photodiode would be sufficient to receive the information.
What's more that wave guide can be fairly long without distorting the injected signal - compared to the size of the computer system even multimode fiber is very good. So it becomes easy to connect chips with 1Gbps (or faster) links. Compare this with todays state of the art - the links between cpus or between cpu and northbridge top out around 1Ghz per line and there is a limit on how many you can have.
If this technology gets developed one can imagine that instead of plugging CPU into the socket one hooks it up to a heatsink, attaches two large wires (power supply) and bunch of fiber-optic links - which go to other cpus, memory, drives, etc.