Comms chips

Startup Demos Optical Processing

One of the first commercial examples of light being used to process signals rather than just carrying them from one place to another was announced earlier this week by Lenslet Labs Ltd.

The Israeli startup says its technology will revolutionize digital signal processors (DSPs) by making them run on optics rather than electronics (see Lenslet Launches Optical DSPs).

Optical DSPs, it says, promise to be thousands of times more powerful than today’s electronic DSPs -- so powerful that they might be used in a much wider range of equipment. In particular, they could replace the ASICs (application-specific integrated circuits) in high performance telecom equipment.

Of course, this isn’t going to happen overnight. So far, Lenslet has demonstrated what it calls a prototype "optical DSP engine." It’s intended to be a replacement for electronic chips (not just DSP chips) in certain situations, according to Avner Goren, Lenslet’s VP of marketing.

Measuring 8x10 centimeters across and just 1.7 cm (0.7 inches) thick, the module is designed to sit comfortably on a board in a standard rack. It's designed to interface to any existing electronics on the board. So, although the guts of the device are optical, its input and output are actually electronic.

The prototype in question is capable of 8 tera operations per second (TOPS) and consumes just 20 Watts, says Lenslet. To get equivalent performance from programmable DSP chips would require 6,600 of them, which would occupy 2.6 square meters of board space and consume 3,300 W, it claims [Read: impossible].

This comparison should be taken with a grain of salt, however. For starters, for serious number-crunching applications, vendors tend to shy away from programmable solutions like DSP chips and go with hardwired ASICs instead.

But Lenslet reckons its technology can outperform such solutions, too. It figures that one of its optical DSP engines can perform the same number of operations per second as 10 ASICs, while consuming roughly one fifth the power, occupying one quarter of the board space, and costing about half the price.

The other reason to be wary is because Lenslet is comparing technologies already on the market with one that isn't likely to be available until 2004, according to company literature.

The numbers will probably do their job, however, in grabbing the attention of potential customers.

Of course, the idea of using optics to make highly-efficient computers is not new. Far from it. It's long been known that optics would be very good at performing calculations in parallel because, unlike electrons, beams of light don't interact.

Also, by its nature, light is intrinsically good at carrying out certain types of calculations that are difficult to break down into simple instructions such as add, subtract, multiply, and so on. An example is a mathematical function called a Fourier transform. When light shines through a lens, the diffraction pattern it creates on the other side is the Fourier transform of the original light beam. (A Fourier transform is a mathematical tool that yields information about how fast a signal is changing.)

So why hadn't this been done before? According to Goren, the stumbling block was the need for fast, multilevel light modulation. People have typically used liquid crystal devices, but, with a top modulation speed of around 200 Hz (cycles per second), these are much too slow for DSP applications, he says.

To get around this problem, Lenslet uses an array of vertical-cavity surface-emitting lasers, or VCSELs, which are modulated directly at rates of 1 GHz. This type of laser is now widely available and used extensively in datacom links (see Laser Blazers).

The VCSELs shine light through a mixture of lenses and mirrors. The resulting output is read by an array of detectors, which sample the signal at 10 GHz.

The optical core is reconfigurable, says Goren. By this he means that different calculations can be carried out by the same unit, either by using a spatial light modulator to gate parts of the optical signal, or by changing the initial conditions applied to the signal electronically. Key parts of the company's intellectual property lie in the electronics that surround the optical core of the processor, he notes.

Goren says that the company plans to release further details of the internal architecture around March 2002.

— Pauline Rigby, Senior Editor, Light Reading
gea 12/4/2012 | 7:44:22 PM
re: Startup Demos Optical Processing A nice little article.
As for why no one every used light to compute with before, I strongly suspect that certain types of image processing "computations" are done by the military in the optical domain. As evidence, consider that the military was almost 3 decades ahead of the outside world when it came to adaptive optics. In addition, a look at some of the optical signal processing textbooks by guys who have done a lot of work for the military indicate that the classified stuff must be very advanced.
BTW, the (spatial) Fourier transform of an image in a lens occurs at the back focal plane of the lens.
opticalwatcher 12/4/2012 | 7:44:18 PM
re: Startup Demos Optical Processing This doesn't look like a 'processor' in the traditional sense. You can perform certain transforms on light with optics, but this is very different than a processor with an adder,
accumulator, and storage.
rustybridge 12/4/2012 | 7:44:15 PM
re: Startup Demos Optical Processing God almighty! Geesh! This is by far NOT a new idea... There are tons of people who have worked on similar things for decades...

Who's that prof at CMU? Dave Casasent, was it? He's been working on using lasers and optical diffraction gratings to so massive amounts of computing using FFT's, etc., for over a decade. I remember hearing him do a colloquium which was amazingly inspirational. By the end of it, everyone is ready to pitch their current work and go into working on that stuff... then you get into it and find all the real problems truly involved in it...

Is it new? No. Does this company deserve amazing amount of attention? Probably not. Is it great PR and a wonderful, fun story that you shouldn't scratch too much below the surface? Probably... Should someone want to invest in such a company when it IPO's? Yeah, sure... probably... if you were the type of investor that jumped on the AI bandwagon with all its promise yet to be delivered decades later. Does it smell like a VC/investor scam? Well... if it smells like a duck...

77thlightguy 12/4/2012 | 7:44:08 PM
re: Startup Demos Optical Processing Lets see, multiple, convolve, multiply, or CMC ... Talk about old stuff, from the dust of time! In the old days it was 850nm laser diodes from Hitachi and multichannel Bragg cells from Crystal Technology. Now it is VCSELs and MEMs and mirrors. Can you spell SBIR phase 1 proposal winner? Good luck if this is how you think you are going to buy that cool custom house in the hills...those 6000 DSP chips will be a lot cheaper...but maybe the new defense budget will bless you.

Sign In