Optical Processing News Story

Lazy8s

Veteran
Israeli Processor Computes at Speed of Light
Wed Oct 29, 5:03 AM ET Add Science - Reuters to My Yahoo!

By Tova Cohen

HERZLIYA, Israel (Reuters) - An Israeli start-up has developed a processor that uses optics instead of silicon, enabling it to compute at the speed of light, the company said.

Lenslet said its processor will enable new capabilities in homeland security and military, multimedia and communications applications.

"Optical processing is a strategic competitive advantage for nations and companies," said Avner Halperin, vice president for business development at Lenslet.

"Processing at the speed of light, you can have safer airports, autonomous military systems, high-definition multimedia broadcast systems and advanced next-generation communications systems."

An optical processor is a digital signal processor (DSP) with an optical accelerator attached to it that enables it to perform functions at very high speeds.

"It is an acceleration of 20 years in the development of digital hardware," Lenslet founder and Chief Executive Officer Aviram Sariel told Reuters.

The processor performs 8 trillion operations per second, equivalent to a super-computer and 1,000 times faster than standard processors, with 256 lasers performing computations at light speed.

It is geared toward such applications as high resolution radar, electronic warfare, luggage screening at airports, video compression, weather forecasting and cellular base stations.

Lenslet said its Enlight processor, unveiled at the MILCOM exhibition in Boston this month, is the first commercially available optical DSP.

"Optics is the future of every information device," said Sariel.

Jim Tully, vice president and chief of research for semiconductors and emerging technologies at Gartner Inc, said most companies working with optics focus on switching optical signals for telecommunications rather than processing information optically.

"I'm not aware of any company that has taken it to the extent of processing optically," he said.

Lenslet has raised $27.5 million so far from such investors as Goldman Sachs, Walden VC, Germany's Star Ventures and Chicago-based JK&B Capital.

PALM PILOT (news - web sites) SIZE

The company's prototype is fairly large and bulky but when Lenslet begins to supply the processor in a few months it will be shrunk to 15 x 15 cm with a height of 1.7 cm, roughly the size of a Palm Pilot.

"In five years we plan to shrink it to a single chip," project manager Asaf Schlezinger said.

Tully said one issue is whether this technology can be produced in volume the way silicon chips are made.

"Because semiconductor manufacturing technology is well developed, you can produce millions at quite low cost," said Tully, who is not familiar with Enlight.

Lenslet said its processor will be competitive in price with a multi DSP board.

Sariel is negotiating joint projects with companies and/or government agencies in the United States, Europe and Japan to produce the processor for specific applications. It already has projects signed with Israel's Defense Ministry.

"We don't rule out licensing our technology to others," Sariel said. "We are looking at a virtual production line where production is done by others and we provide testing equipment."

Tully said semiconductor companies are working on technology that would use optical channels inside a chip to allow very high speed communication from one part of a chip to another.

"It's conceivable this technology could become mainstream inside chips in 10 years time," Tully said.
http://www.reuters.com/newsArticle.jhtml?type=topNews&storyID=3711894

It'll be interesting to see them explore this frontier a bit with their optically accelerated DSP.

In the near future, something along its lines for game processing would be best suited by a specialty market like the arcades. There, legacy considerations aren't an issue, and new architectures are regularly introduced every couple of years with a new platform. The commitment isn't as large there either, so it's easier to take a chance on higher experimental-risk technologies. Initial size estimates of "roughly" that "of a Palm Pilot" for the processor would be better off in the spaciousness of a cabinet than that of a console, a cost that's "competitive in price with a multi DSP board" puts it in line with arcade tech which isn't so cost sensitive, and the expected lower volume manufacturing with this kind of optical processor targets the demand of the arcade market well.

The initial sales pitch of this technology looks to be focused on government agencies for security, defense, and military purposes. This could potentially get it into the hands of government contractors like Lockheed Martin. In the past when designing high-end training simulators, no expense was spared to work with the most advanced processing technologies. Collaborative efforts then with companies like SEGA were used when working with the military simulators to develop the rendering techniques and programming that could use such power. That's how the Model X line of arcade boards came about, and SEGA would still be the best candidate for that kind of collaborative development nowadays. While the arcade sector is dead to most, they are one of the few remaining players and have always been one of the biggest. The market is quite central and profitable to them as they actually derive almost two-thirds of their net sales from it (in contrast, their console software business isn't even posting profits.) And SEGA has been the only arcade company to consistently take chances with such robust boards.

The leap in performance such a thing could potentially provide would put the hardware way ahead of the console curve like SEGA's custom arcade boards always are. Which, of course, means waiting 'till the next generation of consoles after that to get ports of the resulting games that would even approach the originals... an all too familiar scenario! Although, with the right combination of companies all jointly agreeing to bank on these optical DSPs and to put them to use within their product lines, an ambitious production schedule could be achieved for an Xbox3, for example, or maybe even some new SEGA console, if they ever got back into it.

What this Enlight product really represents is good progress, as optics move towards becoming more integrated into microprocessors and a part of the fabrication processes for semiconductors. Convergence with current models will be necessary to achieve mass production, so I think we'll see the spread of optical tech within our established lines rather than seeing current products all replaced by some kind of massive market shift. And we'll get to see what kind of changes all of this power brings to graphics when leading designs like the ones from PowerVR, for example, are adapted with the new manufacturing advances.

Such extra power would be incredible for games like Shenmue where tons of game logic is used to dynamically model the behavior of a whole city's worth of inhabitants.

i_products_enlight.jpg
 
...

What's so good about this technology is that this processor is logically uniprocessor, whereas competing technologies are logically multiprocessors.
 
DeadmeatGA posted this, but the thread was locked.

But this is a pretty good stuff for opto-electronics.
 
What the thing actually does:

http://www.eetimes.com/story/OEG20011008S0024

Optical DSPs promise tera-ops performance

By Patrick Mannion

EE Times
October 8, 2001 (11:59 a.m. ET)

MANHASSET, N.Y. — Startup Lenslet Labs has demonstrated an optically based digital signal processing engine (ODSPE) that has the potential to take DSPs from the current giga-operations-per-second (Gops) limit to tera operations per second (Tops) by 2005, the company announced at the inaugural Communications Design Conference in San Jose last week.
Targeted at wireless (third-generation basestations and fixed wireless access) and wireline (xDSL) communications, along with JPEG, MPEG, machine-vision and voice-recognition applications, the technology combines proprietary encode/decode and signal-conditioning techniques with off-the-shelf optical components. The result is a solution that has the flexibility and programmability of field-programmable gate arrays (FPGAs) and DSPs with the performance of ASICs — without the process limitations of all current silicon technologies.

"Moore's Law [as it applies to conventional silicon processes] is helping designers reach higher and higher integration levels and processor speeds," said Ron Levy, director of marketing and business development at Lenslet (Ramat-Gan, Israel), "but they still can't meet the capacity demands of both wireless and wireline communications systems." All the while, he said, the power consumption is rising on a parallel path with those higher integration levels. "Also, by 2010, the physics of those silicon processes will mean that Moore's Law will likely have run its course," he said. "ODSPE will take processing to a higher level and with a much faster growth curve [than Moore's Law]."

The company has already demonstrated an 8-Tops, 20-watt device. "Using conventional DSPs to get that performance," said Levy, "you would need 6,600 devices, thereby requiring 26,400 cm2 of board space and 3,300 W. FPGAs would need 40 devices and 200 W, while ASICs would need 21 devices and 133 W." As for pricing, Levy claims the ODSPE comes in at 1,000th that of current DSPs, again for the same performance level.

The very nature of the ODSPE's operation, combined with its high processing speed, will also allow designers to operate at a much higher level of abstraction to greatly accelerate the migration from basic research to product development, Levy said.

"Their technology is one of the most exciting developments I've seen in a while," said Will Strauss, president of Forward Concepts (Tempe, Ariz.). "If they can do what they say they can do, they could force everyone to reevaluate the potential of their own technologies going forward."

Lenslet is targeting a market that it believes is one of the fastest-growing in the industry. "The Semiconductor Industry Association predicts that the DSP market will start rising again in 2002 at a rate of 30 percent to reach $6 billion," said Levy. By 2004 the SIA sees it reaching $10 billion. This after a disastrous 2001 period in which the market is projected to drop 26 percent.

Lenslet's core competence is the processing of transforms such as fast Fourier transforms/inverse fast Fourier transforms (FFTs)/(IFFTs), discrete cosine transform (DCT), discrete Fourier transform (DFT), compression, vector-matrix multiplication, equalization and correlation, all of which are essential functions at the baseband level of most communications technologies — "and are the ones that demand the most processing power," said Levy.

Light's transform

However, while designs to date have thrown more DSP/FPGA/ASIC horsepower at the processing of those transforms, Lenslet is instead relying on the fact that as light propagates through optical elements it undergoes a type of transform. ("A simple lens can do an FFT," said Levy.) The input and output data is the light, and the optical elements that perform different mathematical operations on the light represent the linear transform.

The type of transform to be performed is controlled by altering the characteristics of an optical element within the device and also by patented algorithms — specific to each transform — that encode and decode the electrical signals in and out and perform the signal conditioning.

"One of the key differences here between ODSPE and conventional techniques," said Levy, "is that the transform is completed on the whole vector at one shot, and at the speed of light. Conventional processing techniques require the data to be broken down to bit-slice operations or DSP instructions — we operate on functions instead of instructions." That allows operation at a much higher level of abstraction and speeds time-to-market, Levy said. "The technology is also highly scalable, has low power consumption and uses inexpensive optical components," said Levy, namely vertical-cavity surface-emitting lasers (VCSELs), compound lenses, spatial light modulators (SLMs) and optical receivers.

Product pair


Lenslet has outlined two product implementations of its technology. The first is the EnLight256, a general-purpose, off-the-shelf, reconfigurable optical transform engine that supports standard transforms such as correlations, FFTs, DCTs and DSTs. Complementing that is the Customized-EnLight, an application-specific optical transform engine that is available as a component, part of a chip set or as a board-level solution.

To date, forays into optical processing have been limited to optical switches and routers, as well as some attempts at CPUs. "CPUs have a long way to go," said Levy, "and what we're doing is way beyond simple switching and transmission." The closest analogy, said Levy, would be the work done by AT&T on optical correlators. "But they gave up in the mid-80s," he said. "The algorithms were too complex."

Attempts to perform optical transforms specifically have been limited by what Levy classifies as the I/O barrier. "The problem was in the conversion of the electrical signal to multilevel optical signal [light modulation] and the reconversion to electrical after the transform was completed," said Levy.

"The speed of the typical liquid-crystal modulator [which is fed by a single light source and requires physical changes in the LC elements to produce 8-bit gray-scale output] tops out at 200 Hz — much too slow for DSP applications."

Lenslet's approach uses VCSELs that are driven by an encoded signal that produces a digitally modulated, continuous analog waveform. The VCSELs have a 1-GHz frame rate.

That optical output then passes through a series of compound lenses and a specially adapted SLM. The characteristics of the latter determine the nature of the transform to be performed. Those characteristics can be changed in microseconds, Levy said, and thus allow on-the-fly programmability for such applications as software-defined radios.

The light exiting the SLM is then put through another series of compound lenses and detected by a photodetector array operating at a 10-GHz frame rate. The technique uses free-space optical transmission with no proprietary channeling or lightwave transmission techniques.

While the ODSPE takes care of the transforms, other DSPs will be required for lower-level processing for both the encode/decode algorithms and other baseband functions of future designs. "These can all be low-cost, low-performance devices, as the core processing has already been done by the ODSPE," said Levy. "So, the cost of support chips is greatly reduced."

While the company is demonstrating product now, it expects prototypes to be available next year. "These will mostly [be] used for evaluation and feedback, but by 2004 we expect to be shipping product for designs entering the market in 2004," said Levy. "The important parts were the breakthroughs in the encode/decode algorithms. We're now working on modified algorithms for the various transforms to be performed."

In a nutshell:

It does 8Tflops - but only for Fast Fourier Transforms and other standard DSP operations. It still uses a lot a silicon-based logic to support the optical transform core. I'd be hard-pressed to find a gaming application for this.

At the risk of sounding self-promoting, here's an analogy I wrote in a thread that was subsequently locked.

Not to say that this isn't a impressive piece of technology - it is (opto-electronics is a popular field right now) - but this is fundamentally different from CELL or any other CPU. It's like comparing a ski lift to car. A ski lift is very good at what it does, and maybe in people-meters/sec it is very impressive, but it is very limited. A car is not as good for taking people up slopes, but it can do a lot more.
 
I wonder what the precision and dynamic range is...

It's very interesting ( almost reminds me of some of the early analogue computers and hybrid machines that drove differentiator/ integrater networks
via dacs to solve differential equations.. )

They quote TOPs not TFlops.. as well ( bytewise SAD functions ( used for motion search ) can peak at 32 ops/cycle )
 
Lazy8s - I pretty much agree with you about how SEGA's Model X series of arcade boards came about. they were based on advanced military simulation technology. It was not really experimental or radical technology by any means. It was simply some of the strongest and most robust rendering technology avaliable. which had been in use for 5-10 years (texture mapping) or even decades (polygons)


As neat as this opitical technology might be, I'm sure it wont be used by any arcade manufacturers like Sega or Namco, for the reason that, its not the right sort of technology. but eventually, something will come along that sparks the resurgence of arcade games.
 
Why is this old dusty thing getting resurrected?

Reading through the thread, it's pretty clear this has nothing to do with console gaming.
 
It's a new technology (at least this implementation) that promises to make the type of calculations needed for 3d graphics (vector matrix transformations) extremely fast at a very low price.
If that is not related to videogames and consols, I don't know what is.

As I've written before, I can see no possible harm done in bringing up an old thread, if you think it's relevant.


Edit: Denmarks national radio, conducted an interview with a guy who works at Lenslet, it's not too boring, so if you can stand to listen to a little danish, then it's 2:55 minutes in the feature, interupted for a few of seconds now and then by the host.
javascript:
Click on adress: http://www.dr.dk/Videnskab/Harddisken/forside.htm?programmer-stat=true

And then cut and paste the below to your browsers address box:
javascript:DRPlayer('','','/NR/rdonlyres/C4531CD6-5D9A-4195-A574-CACE2533AF2E/5565/e5ecc5f51c324e82bea90d2a560950ce.asx','')
 
Back
Top