Slashdot Mirror
UCLA Researchers Demo First Silicon Laser
An anonymous reader submits "Researchers at University of California, LA have demonstrated the first silicon laser. The lack of a silicon laser has been a major roadblock in the progress of silicon optoelectronics and photonics. This development shows that despite popular belief, a laser can indeed be made on a silicon chip. Modern electronic computers are getting closer to being optical in any case (gigahertz range). This discovery makes optical computers much closer to reality."
No, they aren't made of silicon. They're usually made of compound semiconductors, such as gallium arsenide, etc.
Ce n'est pas un vrai mouvement de robot!
This is nowhere near practical for integrated photonics. For one, the repetition rate is limited by the free carrier lifetime, which can range anywhere from about a nanosecond to several hundred nanoseconds, depending on how the waveguides are fabricated. So, the fastest the pulses can be repeated is about 1 GHz. Which is just slow even by todays standards. In this work they didn't even do that good, they repeated it at 25 MHZ. Slooow! And second, and the biggest problem, the Raman effect relies on an external laser to pump the Silicon laser! These can easily cost $25000 dollars! Using the raman effect there is NO way to do electric pumping. Another external laser will always be necessary, so I don't ever see this becoming practical.
And at this point there is no other easy way to make a silicon laser using anything other than the raman effect. Any other method would require the use of exotic materials (i.e. erbium, heterojunctions, etc.). It's been tried and the results aren't all that promising.
Regardless, this is some great progress and with some more research it can be improved on. For one, they are putting the silicon chip in a fiber ring to form the laser. The next obvious step is to make the laser work just on the chip. One easy way to do this would be to put mirrors on both sides of the chip. You can easily make mirrors with some fairly simple fabrication.
Two problems arise. Driving a signal from one end of the chip to the other end is very slow because the wires present a high RC load to the puny transistor. The other problem is simply routing the wires.
The laser solves the first problem because we can simply transmit a bit (0 or 1) by modulating the laser light. The bit will travel at the speed of light.
The laser also solves the second problem because there is no need to route the optical paths. The light from one laser can cross the path of light emanating from a second laser and can continue, unimpeded, to the detector intended for the first laser.
A laser on silicon essentially solves the wiring problem of traditional digital integrated circuits (ICs). Modern digital ICs consist mostly of wires; a small percentage of the silicon area is the transistors that perform the computation.
Not true last I checked, and I am presently doing IC design. Yes, you need plenty of space for your routing channels, but especially with the number of metal layers we have now, there's plenty of space _above_ the active circuitry even after you take out the layers used for routing within the gates.
Two problems arise. Driving a signal from one end of the chip to the other end is very slow because the wires present a high RC load to the puny transistor. The other problem is simply routing the wires.
The RC load problem is only _severely_ nasty if you lay out the bus as one long wire. If you segment it and put repeaters, time scales as length, not as length squared. You can also pack adjacent data stages more closely together as linewidth decreases, so chip shrinks continue speeding things up.
Routing is only a crippling problem if you're trying to do an anywhere-to-anywhere communications mesh. Even then, there are well-known topologies that minimize pain. This is part of why the movement to multi-core chips is happening - a massively superscalar chip _does_ have to route results from any part of the chip to functional units at any other part. Smaller cores in parallel are a win/win scenario (simpler design, easier communications).
See above re. routing. Routing is actually _easier_ now than it was a decade ago, because we have six or more metal layers to do it in rather than two or three. For a pipelined data flow, the routing complexity between stages doesn't really _increase_ with chip size. It's aggressively superscalar designs that do that, and those reached their peak complexity a few years back (with emphasis shifting to multi-core now).
The laser solves the first problem because we can simply transmit a bit (0 or 1) by modulating the laser light. The bit will travel at the speed of light.
Electrical signals travel somewhere between 10% and 50% C, usually. The real problem is capacitive loads, and an electro-optic system has to deal with those too, believe me. Electro-optic systems will actually have it _worse_, because you need large photosensitive structures to pick up the light (more capacitance), and have relatively small photocurrents to charge or discharge them with. You no longer have electric networks with huge fan-out (and large capacitance), but if you build a segmented hierarchical scheme, you don't with an electrical implementation either. And you get an _optical_ fan-out in return - you're only producing so much light, and it has to be split among N potential receivers.
Light is very far from being a magic bullet for intra-chip communication.
The laser also solves the second problem because there is no need to route the optical paths. The light from one laser can cross the path of light emanating from a second laser and can continue, unimpeded, to the detector intended for the first laser.
This turns out not to be the case. Your emitters are small enough that diffraction will spread out the beams quite a bit. So, you're either using a broadcast system (which reduces your communications bandwidth drastically - you only have one channel shared among all components), or you build waveguides. Waveguides have to be routed, and have the same kind of crossover problems wires have. They're actually a bit worse, as they have more constraints on geometry than a low-current signal wire does.
In summary, I don't think that optical communication works as well as you think it does, and the problems you note with electrical communication are largely solved.