Using Lasers to Speed Computer Data

Carl Bialik from WSJ writes "The start-up Lightfleet has developed an unusual way to use lasers to speed the flow of data inside a computer, hoping to break a bottleneck that can hamper machines using many microprocessors, the Wall Street Journal reports. The company plans to sell servers it predicts will be much more efficient than existing systems in tackling tough computing problems. Tasks could include automatically recognizing a face in a video image or sifting through billions of financial transactions for signs of illegal activity. These machines will attempt to sidestep some of the problems associated with parallel computation by ensuring all processors are connected, all the time."

Not really a new idea

[twiddlingbits]

Nohing new to see here..move along...the TI chip that is in the DLP TVs does this already. (from wikipedia) In DLP projectors, the image is created by microscopically small mirrors laid out in a matrix on a semiconductor chip, known as a Digital Micromirror Device (DMD). Each mirror represents one pixel in the projected image. The number of mirrors corresponds to the resolution of the projected image. 800x600, 1024x768, 1280x720, and 1920x1080 (HDTV) matrices are some common DMD sizes. These mirrors can be repositioned rapidly to reflect light either through the lens or on to a heatsink (called a light dump in Barco terminology). The rapid repositioning of the mirrors (essentially switching between 'on' and 'off') allows the DMD to vary the intensity of the light being reflected out through the lens, creating shades of grey in addition to white (mirror in 'on' position) and black (mirror in 'off' position). This system is just a new application of a technology invented in the late 1980's. No reason it's groundbreaking and no reason it shouldn't work in theory. Solid state lasers are very reliable and have a long life time. However, I don't know of any chips that have the ability to directly receive laser light pulses from a source and convert them to 1's and 0's. And keeping the lasers, mirrors and receivers aligned might be tough.

Re:hmm

[mikael]

The mirror will be fixed in place, so that shouldn't be a problem.

Normal interprocessor communication would require a crossbar switch or some kind of virtual network to support different grid configurations (square grid, cube mesh, torus, or hypertorus). Usually each node has a router to handle this for it. This system gets rid of the routing and just multicasts each packet of information.

Presumably receiving a data packet prevents a node from sending out another data packet at the same time. Although, this would seem to make the system act as one serial communication line. The benefits of having multiple connections is that messages can be sent between nodes in parallel. Occam had the concept of North, East, South and West connections.

Re:o_0

[I_HATE_THIS]

Use different wavelength, CDMA in a beam. And I dont think line-of-sight is necessary with fiber optic cable. Sound like there is no switching, so, it is up to the local processor filter out unwanted messages (sound like Tibrv with light)