Slashdot Mirror
Researchers Develop Photonic Processors
TheCybernator writes to mention a New Scientist story about scientists who are developing a light-based processor by actually storing and delaying photons. These 'optical buffers' may one day be used to make super-fast microchips based on light instead of electrons. From the article: "A decade from now ... there [may] be not seven cores but hundreds on a chip ... Connecting these cores using light could solve this problem. Until now, the lack of optical buffers has been a key roadblock to these kinds of light connections. The way information is transmitted means that buffers must hold packets of data while a router decides where they are to be sent. Buffers are also needed to delay optical pulses - so they do not collide at switching points - and to synchronise streams of data coming from different places."
Something like 20 years ago, I heard about a device that used interference in a crystal with the upshot that the presence of light on one facet of a crystal determined whether light could go through the crystal along another path. This was supposed to form the basis for entirely-optical logic devices. Does this ring a bell? Anyone?
-jcr
The only title of honor that a tyrant can grant is "Enemy of the State."
The world market for photonic computers will only be five or six.
The important distinguision to make when comparing the benefits of going massivly paralel processing is that it is possible to solve NEW problems in realtime with these processors. Eg, we don't need to run Word 100 times faster, however we can get eg. games and scientific simulations (two sides to the same coin) that uses detailed physics engines and realtime raytracing. Raytracing can be almost naivly paralellised with up to as many processors as screen pixels. I remember using a computer with 65536 processors called the maspar which was built in the early 90's. Our main use for this computer was for image processing which also could easily be parallelized. It just took a bit of a shift of perspective to learn how to program it since it was SIMD (Single Instruction Multiple Data) but boy where it fast for it's time.
Physics is a bit more difficult but there are tehniques too for paralellization utilizing the fact that object interactions form islands of connected parts. Eg, when simulating your hair in a realistic way don't test for interactions with the objects in a distant part of the game. Physics engine are just starting to become used for these purposes but can easily require how much CPU power you want for it. Simulating eg. the clothes in the game characters or dynamic subdivision of parts as they break or bend due to forces (do you want realistic dents in your car after hitting that pedestrian?). These would both require an order of magnitude more CPU power than what we can do in realtime physics today.
So, to make a short summary. Yes, we can always achieve new tricks with even more computing power. Give me a cluster of a million processors and i would still complain that it's too slow for what i want to do.
We are starting to get to the point where the capacitance of the tiny little wires in a genuine concern, and crosstalk between them is also significant. Also, the amount of space taken up by wiring is annoying. You can use a single waveguide with several frequencies of light to replace several wires and solve all those problems at one. At least, in theory. In practice, it's really hard to build it. But, it'll be pretty sweet when we get it all sorted out.