Driven by ever more complex algorithms and more and more channels for the processor to handle, the need for more digital signal processing speed is escalating rapidly. Conventional DSPs are pushing the silicon technology envelope to meet the challenge. They are following both expected and unexpected strategies: Increasing clock speeds—Texas Instruments recently announced a 1 GHz product; and, adding hard wired and configurable coprocessors to execute specific algorithms.
The stress of keeping pace with the demands of compute-hungry applications is only getting more severe, however, which has led some engineers and venture capitalists to consider introducing a technological discontinuity. In this case, it is light-wave processing.
The concept of optical digital signal processing in some ways is a misnomer because optical signals are by their very nature analog. Nevertheless, there is a rationale for using the term—the devices are best suited for the same applications as DSPs. But employing massively parallel streams of protons traveling at the speed of light can—theoretically at least—yield processors that are orders of magnitude faster than even the fastest silicon DSP.
Abundant Applications
While there are plenty of challenges to be overcome, optical digital signal processing (ODSP) would be a good fit for a number of compute-intensive applications where the physical size of the system is not critical. Among these applications are:
- Interference canceling between cellular users (multi-channel distortion)
- Software-defined radio
- Real-time video compression for H.264 and HDTV
- Homeland security in applications such as voice analysis and face recognition
- Radar and electronic warfare
- Correlating gene data in biotechnology
- Data mining of large networks.
Conceptually, optical digital signal processing hinges on the observation that although DSPs handle many different types of calculations, they are most useful for executing algorithms such as FFTs. As it turns out, light can undergo similar transformations when it passes through optical elements.
 |
| |
SLMs
Spatial light modulators are large 2-D arrays of densely packed light modulators. The number of individual modulators can range from one thousand to more than a million. They are usually electrically addressed. Light modulation in the form of intensity, phase, or polarization is controlled by applying different electrical signals to each cell. Liquid crystals and magneto-optic materials can be used to construct SLMs but their shortest response times are typically of the order of 10-100 µs. To achieve nanonsecond switching, multiple quantum level (MQL) SLMs fabricated in GaAs have been introduced.
|
|
 |
In particular, devices called spatial light modulators (SLMs) can alter the light to match the most intensive aspects of the numerical manipulations required by the algorithm.
This type of technology has, in fact, been around for quite some time. But it has also been limited by internal switching speed considerations (electronic switching and lens motion) and accuracy (or resolution) considerations. It also had a problem with what might be called being "hard wired" in an optical sense: The SLMs could handle only one algorithm and only one set of parameters for the algorithm. Programmability of the SLMs is critical for commercial use, of course, and that is one of the key technology challenges.
Not surprisingly, semiconductor vendors have investigated light processing. Texas Instruments, in particular, has for some time had a successful commercial operation based on micro-mirror technology. TI's Digital Light Processing (DLP) operation, however, is primarily focused on displays that range in size from readily portable projectors to systems that can be used to display motion pictures on theater screens. More information can be found on the unit's Web site at www.dlp.com.
TI has not aggressively pursued a commercial implementation of an ODSP itself, says DLP Business Development Manager Wes Stalcup. He cites several reasons, including the speed penalty that must be overcome when light energy is covered back to electrical energy so it can be used by the rest of the system. Projected markets for a system that might be three feet square also figured into the analysis.
TI also concluded that a board populated with an array of its 1 GHz, TMS320C64xx chips could match the speed of any optical products that might come on the market in the foreseeable future—and be more in line with TI's product strategy.
But that does not mean there is no interest in ODSP at TI. The DLP's operations DMD Discovery Kit provides developers with the ability to use micro-mirror technology to manipulate the carrier beam itself (attenuation, shape or refine the outbound optical signal) or, use micro-mirrors and conventional DSPs to manipulate the information on the beam (such as multiplication) using micro-mirrors instead of SLMs to modify the light.
Making It Work
On the other hand, speed is the major advantage optical digital signal processing offers the industry. Since light is relatively immune to crosstalk and other noise inducing phenomenon, light processing can easily take advantage of massive parallelism.
|
| |
Compound Lenses
Compound lenses focus light from VCSELs on the SLM pixels and focus light that passes through the SLM on the photo detector array. They are called compound lenses because two lenses are used to achieve superior resolution. The same lenses used in CD-ROM players can often accommodate the requirements of SLMs.
|
|
|
Whereas conventional electronics typically requires multiple manipulations of data arranged in series, optical processing can easily perform the transform in a single clock cycle—which equates optically to a single pass of light through the SLM and the other system components. All of the components—vertical cavity surface-emitting lasers, compound lenses, and the SLMs—are relatively inexpensive.
But optical processing is useful only for the vector multiplication segments of algorithms. Other processing functions associated with silicon DSPs—such as encoding and decoding—must be accomplished with conventional electronic circuits.
Precision is also an issue. Light information is analog which means it's infinitely precise until it is converted into electrical energy. But as soon as it is quantized, some level—a significant level—of precision is lost because the conversion technologies are non-linear.
The technology as practiced by one ODLP company can be explained in terms of a simple vector matrix multiplication. In this example, the value being measured is light intensity. Resolution is 8 bits for reasons mentioned in the previous paragraph. Referring to Figure 1, a linear array of 256 VCSELs (vertical cavity surface emitting lasers) provides the input values for the vector multiplication. The processor is a 256-by-256 SLM (spatial light modulator.) A linear array of 256 photodiodes detects the result and 256 ADCs convert the analog result from the photodiodes into the digital domain.
Figure 1: Vector matrix multiplication (Courtesy of Lenslet) |
|
| |
VCSELs
Vertical cavity surface emitting lasers are used as the light source for matrix multiplication. VCSELs are specialized laser diodes that have the robustness of LEDs but emit coherent energy perpendicular to the boundaries between the semiconductor layers. Ordinary laser diodes emit coherent energy horizontal to the layers. Because VCSELs (pronounced vixels) produce coherent light in the vertical plane less electrical energy is required to produce a given amount of coherent light. Almost as important is the fact that VCSELs are relatively inexpensive to manufacture, package, align, and test. They also have lower power dissipation and higher reliability than ordinary laser diodes.
|
|
|
Each element from the 256 VCSELs is projected on a column (256 light modulators) of the SLM matrix. Each row of the matrix is projected on a single photodiode. Remembering that multiplication is the sum of a single value repeated multiple times, it can be seen that the resultant energy in a detector is the sum of the 256 inputs from the row that projected onto it.
In this example, the SLM is acting simply as a means of accumulating values. But SLMs can also alter the information they receive. By changing values stored in the SLM matrix is equivalent to creating a new transform. Only when matrix values can be changed easily does optical light processing stand a chance of becoming a viable commercial technology.
Into the Future with ODSP
An Israeli company, Lenslet, has taken a major step toward establishing a market for ODSPs with the introduction in 2003 of the EnLight256, one of the first commercial devices of its kind to become available. EnLight256 is a fixed point processor that can execute 8000 gigaMACs per second—roughly three orders of magnitude faster than traditional DSPs. Lenslet is not the only commercial vendor. Essex in Columbia, MD, also has products in the ODSP space although it is focused on military applications.
Whenever possible, Lenslet has used off-the-shelf components. It uses optics developed for CD-ROM technology and VCSELs that are widely used in optical fiber communications systems. Even the GaAs chips used in the SLM are created using a fables semiconductor design methodology.
Enlight256 has a standard board-level form factor. It includes conventional memory and I/O as well as program memory and buffers. Most interesting, of course, is the processing architecture, which has three levels that match the requirements of executing algorithms—multiplication, scalar operations, and vector operations.
- The optical vector matrix multiplier (VMM) is the core of the EnLight256 processor. It has already been described as being composed of an array of VCSELs for input data, an SLM for processing, and an array of photo detectors plus ADCs for the output phase.
|
| |
Photo-Detector Arrays
Photo-detector arrays in the form of photodiodes receive the light from the SLM and convert it to electrical energy. Each photo detector is coupled to an 8-bit ADC to complete the dataflow through the optical processor.
|
|
|
The SLM, in particular, is worthy of note. It provides its superior performance because it uses multiple quantum well (MQW) GaAs technology. Applying an electric field across each of the light modulators shifts the energy levels in the semiconductor. This in turn changes the transition energies of the light that is passing through the modulator, which is the basis for making calculations. Lenslet's MQW SLM can switch in nanoseconds. Its high contrast ratio provides 8-bit resolution—or, 256 distinct "gray levels" of contrast. It has a frequency response exceeding 1 GHz and operates at less than 5V. Finally, EnLight256 crams tens of thousands of pixels into a small area.
Functionally, the optical core does just one mathematical operation: multiplication of a matrix by a vector. It runs at a rate of 125 MHz for a multiplication of a 256 element vector by a 256x256 matrix, which cranks out 8000 Giga operations per second. Matrix-vector multiplication can be used to calculate any linear transform.
- For vector-vector operations, Lenslet includes in the EnLight256 silicon a vector processing unit (VPU) that does operations like adding two vectors element by element, subtracting, and finding the maximum value of a vector. The VPU is presently implemented on an FPGA and is a capable processing engine performing about 128 Giga operations per second. It will eventually be integrated onto an ASIC.
- A 32-bit standard, off-the-shelf DSP is included for scalar operations and does all the logical functions, runs the software, and manages board operations.
Software Development
The EnLight256 is programmable, which is one of the reasons it can be considered the first commercial optical digital signal processor product. It has its own instruction set that supports vector-vector and vector-matrix computation. Since the MMU has an 8-bit word length and the VPU has a 16-bit word length, the instruction set supports arithmetic and logical operations for both word lengths as well as real and complex vectors. Also included in the instruction set are vector comparisons, shifts, and other manipulations.
The software development flow starts with The MathWorks' MATLAB and takes those results into EnLight-specific software suites, including a simulator that links the development code to EnLight Function libraries (FFTs, FIRs, correlators, and so on) as well as user-defined libraries. The last step is generating DSP code, which is accomplished by EnLight Studio Emulator.
The results of optical processing are impressive, as shown in the benchmarks provided by Lenslet and shown in Table 1.
| Function |
Lenslet EnLight256 @ 125 MHz (8000 GMAC) |
State-of-the-Art DSP @ 1 GHz (8 GMAC) |
Ratio |
| FIR (1000 samples 128 complex filters 128 taps) |
Cycles |
1000 |
8,200,000 |
x8200 |
| Time |
8 µSec. |
8.2 mSec. |
x1000 |
| Correlator (128 complex of length 128) |
Cycles |
128 |
1,050,000 |
x8200 |
| Time |
1 µSec. |
1.05 mSec. |
x1000 |
| FFT/DFT (128 complex) |
Cycles |
1 |
400 |
x400 |
| Time |
8 nSec. |
400 nSec. |
x50 |
| FFT/DFT (16K samples complex) |
Cycles |
256 |
100,000 |
x400 |
| Time |
2 µSec. |
100 µSec. |
x50 |
| FFT (32K samples) |
Cycles |
512 |
200,000 |
x400 |
| Time |
4.1 µSec. |
200 µSec. |
x50 |
| FFT (64K samples) |
Cycles |
1024 |
450,000 |
x440 |
| Time |
8.2 µSec. |
450 µSec. |
x54 |
Table 1: EnLight256 benchmarks
Size, Yield, and Cost Issues
Although there are many applications for ODSP technology, Lenslet is expecting a gradual ramp that will be determined by factors other than raw performance. Due to the optics, EnLight256 and its successors will be board-level products for the foreseeable future. Within the next three years, the form factor of the board will shrink to about 2 feet by two feet, which makes size a factor in determining suitable applications.
Similarly, over the next three years production volumes are expected to reach into the thousands or tens of thousands, in part because of market acceptance and in part because of the inevitable manufacturability and yield issues every new technology must face.
The EnLight256 has, however, been designed to be robust. Mean time between failures is expected to be close to 10 years thanks to built-in redundancy. The system incorporates 288 VCSELs, for example, instead of the 256 required for normal operations. The system can automatically switch on the fly from a VCSEL that fails while the system is in use to one of the 32 extra lasers.
Lenslet also expects to meet aggressive pricing goals, especially considering the orders of magnitude higher performance the technology provides compared to silicon. Vice President of Business Development Avner Halperin says the EnLight256 will be price competitive with leading, high-performance multi-DSP boards.
Technology discontinuities of the type that Lenslet envisions occur relatively infrequently but when they do everything around them changes as well. The most striking technology discontinuity in the past 50 years, for example, is the replacement of manual calculation methods with electronic in the form of microprocessors, microcontrollers, DSPs and other processors. While there seems little doubt that as semiconductor technology starts to reach theoretical limits lightwave technology will play a bigger role, no one knows when it might happen.
About the Author
Contributing writer Jack Shandle is a former chief editor of both Electronic Design magazine and ChipCenter.com. He holds a BSEE degree and has written hundreds of articles on all aspects of the electronics OEM industry. Jack is president of eContentWorks, a consultancy that creates high-value content for publishers, eOEM corporations, and industry associations. His email address is
jshandle@earthlink.net.
