New Optical Chip Claims 8 Trillion Operations/sec.

Richard Finney writes "Lenslet is announcing the 'World's First Commercial Optical Processor.'. Reuters has the story here. The Inquirer has a cool graphic here on it. The processor is specified to run at a speed of 8 Tera (8,000 Giga) operations per second, one thousand times faster than any known DSP. When Lenslet releases its Enlight processor in a matter of weeks, a unit using the technology will be 1.7 centimetres high and measure 15 by 15 centimetres."

Comments

FYI

[Doesn't_Comment_Code]

Interstingly, optical processors aren't faster because light is faster than electricity. They are faster because they have much faster rise and fall times between digital on and digital off.

Re:FYI

[QuantumFTL]

Interstingly, optical processors aren't faster because light is faster than electricity. They are faster because they have much faster rise and fall times between digital on and digital off.

While your statement may indeed be correct, that is not why this chip is faster. The reason is that they are doing analog signal processing using the physics of various optical elements to perform computationally intensive mathematics.

Think of it this way: We can use large, expensive mathematical operations to simulate optical components, which means we can also do the reverse - using optical components to perform the expensive mathematical operations.

I read about this about 2 years ago, and it was really quite fascinating to me. It turns out that with a simple lense, you can compute a fourier transform just by focussing the light (it doesn't focus down to an infinitesimally small point).

I managed to find an article about this, hopefully it should be apparent why this chip doesn't run quake:
Check it out here.

They are certainly not the only people doing this. I've seen plenty of references of this being used in missile guidance systems (turns out a simple fourier transform trick can be used to track objects in a camera). Someone I met while working at the Jet Propulsion Lab was working on this Optical Signal Processors. They prove to be very big in the next 10 years.

Cheers,
Justin

Re:This is the Future

[back_pages]

For real, what's the point of 8 trillion operations per second when there's no existing memory to support it (of which I am aware)? So this chip runs REALLY REALLY FAST on code that's REALLY REALLY SMALL, and otherwise it's bottlenecked by the memory bus and memory speeds.

I appreciate that it's a great demonstration of new technology, but maybe it's a little premature to call this a new commercial chip. It sounds to me like a demonstration of a research project or an exposition of things to come.

It's quite possible that I'm completely ignorant about this, but to whom do they expect to sell the latest and greatest THREE ORDERS OF MAGNITUDE increase in memory bottleneck?

Memory is irrelevant for this kind of "processor"

[wowbagger]

This is NOT a Harvard architecture part - this isn't fetching instructions from RAM and executing them, like a regular DSP would.

Think of this more like an FPGA - you have a device that is configured for a specific processing algorithm, and data is fed in at wire rate and processed at wire rate.

An example of how a device like this might be used may be in order:
I'm trying to find a radar pulse buried in the noise coming in from my receiver. I want to know the phase delay of the radar pulse - how long from when I sent it till I got it back.

Now, I know what my radar pulse looks like as it goes out. I know that any reflection is going to consist of versions of that pulse shape, delayed and of varying strengths. So what I do is called a correlation - the easiest way to think of this is to imagine having 2 transparencies, one of my outgoing pulse, and one of the incoming signal. Now, I hold them up to the light, and slide the incoming signal across the reference pulse until things match up - that's the point of maximum correlation, and that give me the delay of the signal.

A real correlation function is a bit more complicated as you have to allow for the signal level to be changed - if I am looking for a signal of N samples in a received data stream of M samples, I have to do M*N multiply and add operations to get my correlation. Now, for a radar signal I might be sampling at over a billion samples a second, and looking for a chirp of a 100 ns would give me over 100 billion MAC operations a second. There are ways to do that with conventional DSPs, but they are a galloping BITCH to do (you basically make a cluster of DSPs, and each DSP takes a part of the signal. Synchronising that is a bitch.)

This device would work by having the shape of the outbound pulse represented in the structure of the device itself, and the MACs are done by taking the incoming data stream and projecting it on the structure - thus you do all your processing in parallel, and at wire speed. You get a pulse out when the incoming signal matched the signal you ar looking for.

Re:More Info

[Mozz Alimoz]

I'll have a guess at how this works.

The article says "The Ablaze(TM) is the Spatial Light Modulator (SLM) in the optical core of the EnLight256(TM)". Going by the graphic in the Inquirer article, they shine a row of blinking lights through a LCD-like device (and some lenses and mirrors I assume) and collect the results in a column of light sensors on the other end.

Each pattern of on/off elements on the LCD-like device gives them a different transformation running at however fast you could emit and sense the light. I doubt they mechanically move the optical arrangment so that would seem to limit the number of transformations. Some of the LCD patterns might give useful transformations. A vector multiply, a Fast Fourier Transform (maybe) or a sort (I doubt it)?

If the numbers are an analog light intensity level the precision would depend on how precise the light emitters and sensors you have are. Packaging the mirrors and lenses small enough is a neat trick. Having a problem that fits the available transformations and can supply data in and out fast enougth is another. I wonder anything useful can be done by quickly switching LCD matrix pattern, or directly feeding outputs back as inputs?