By making analog circuits easier to design and simulate, researchers at the Georgia Institute of Technology hope to further the cause of analog while enabling engineers to incorporate lower-power circuitry into their products. In the same way that field- programmable gate arrays have spawned advances in digital signal processing, the Georgia Tech researchers believe their large-scale field-programmable analog array (FPAA) has the potential to seed growth in analog.

Current versions of these chips are unlikely to make it into any but the most demanding specialist applications, such as simulating the neural signal processing done by biological organisms. Nevertheless, they make it possible to design, prototype and test systems quickly and easily, without having to fabricate new chips.

Moreover, FPAA technology allows a wide range of less-skilled users to try out sophisticated, low-power techniques developed by the analog signal-processing community, said Paul Hasler, a professor at Georgia Tech (Atlanta). "Rough estimates suggest there are around 3,000 analog engineers in the world," said Hasler, who has been focusing on the application of FPAA technology. "Compare that with the number of system designers, those working in DSP, etc.; even the most conservative numbers are above a million."

If FPAAs encouraged even a small proportion of those engineers to start using analog technology, it would make a huge difference. "The aim is to feed these approaches into mainstream engineering," Hasler said. "That process is starting by providing a strong educational infrastructure, which is already happening at the grass-roots level."

Analog signal processing will not overtake digital but will complement it, Hasler said. He's working on a new way of thinking about mixed-signal processor design--cooperative analog-digital signal processing--that gets the best out of both technologies. Analog preprocessing can ease the A/D bottleneck "and reduce the computational workload on the DSP that follows," Hasler said.

Eugenio Culurciello, an assistant professor at Yale University, thinks FPAAs "will advance the integration of programmable analog blocks in a way similar to FPGA and digital circuits. They have enormous potential for reducing the prototyping time for small analog circuit assembles . . . and should also have the ability to address large-scale arrays." FPAAs, he said, "can provide adequate performance for most analog applications and even for integrated instrumentation."

Analog signal processing
Although analog circuits are not ideal for traditional symbolic computation, they have many advantages for processing signals coming in from sensors. First, there is no requirement for an analog-to-digital conversion step, which eliminates the need to make arbitrary decisions about which information to keep and which to throw away. This applies to both data resolution and timing.

In addition, with analog computations, the scaling between resolution and computation time is often much better, because each additional "bit" does not require its own operations.

For combined sensor/processor chips, such as imaging arrays, there are also the advantages of parallelism and retaining geometrical integrity. In devices like the cellular neural network, all nearest-neighbor interactions in an array happen simultaneously. This means that an entire acquired image can be processed in relatively few steps. Experiments have shown, for instance, that a Laplace transform can be performed up to three orders of magnitude faster in the analog domain than in the digital.

Finally, 20 years ago, Carver Mead, then at California Institute of Technology, showed that circuits could process signals with much lower power if they were not forced to operate in digital mode. That insight came from Mead's noting the efficiency of biological neural circuitry, which is much more analog than digital in operation (it uses the physics of the neuron to compute, rather than some step-by-step algorithm). Neuromorphic engineering, the development of highly efficient circuits based on biological models, is based on this idea.

"Gene's Law" (named for TI fellow Gene Frantz, who first identified the phenomenon) holds that the power required for a million multiply-accumulate operations per second sees roughly a factor-of-two improvement every 18 months. But while digital technology is improving all the time, the power-efficiency gap between it and analog is large. For engineers working on computationally intensive but power-restricted applications, this gap represents a 20-year head start.

But for all but the most demanding applications, analog's advantages have been outweighed by its practical difficulties. Analog design is difficult; testing is time-consuming, because it involves fabrication; and the solutions produced have mainly been very specific to individual problems, because of their lack of programmability. These three problems have combined to produce a fourth: unpopularity among electrical engineering students at all levels.

With the increasing ease and instant gratification provided by digital design, analog has understandably taken a backseat. That has led to a shortage of analog engineers, further decreasing the likelihood that analog solutions would even be suggested, never mind adopted.

An easier way
The technology being developed by Hasler and his colleague Chris Twigg, now at the State University of New York, Binghamton, may eventually solve several of these problems. The FPAAs themselves are networks of analog elements connected using floating-gate transistors--devices that allow the coupling between elements to be strengthened or weakened continuously. Once programmed, they are not just passive wires but are actively involved in the analog computation. In contrast, FPGAs use suboptimal interconnects as a necessary overhead to provide programmability. The long transmission lines that are often required degrade performance without contributing to the computation itself.

The existing devices, with around 50,000 analog elements, should prove useful "in those applications that require a substantial amount of signal processing, have a very tight power budget and need to be rapidly developed and deployed," said Brad Minch, an associate professor of electrical and computer engineering at Franklin W. Olin College of Engineering (Needham, Mass.). "These FPAAs also offer the possibility of customizing or tailoring a product for the end user."

Minch pointed to applications such as "providing speech recognition in cell phones and mobile-computing platforms, and implementing direction selectivity and noise reduction in hearing aids" as examples.

"The main competing technologies in such applications would be DSP and full-custom application-specific integrated circuits," Minch said. "The main advantage of FPAAs over the DSP solution is that the power consumption of the entire FPAA could be lower than that of the analog-to-digital converter module by itself, to say nothing of the power consumed by the digital circuitry." The main advantage over the full-custom ASIC is flexibility, Minch said, along with a faster "design, development and deployment cycle."

Even in applications where ASICs are more efficient, the FPAAs may have the benefit of being more predictable. It's always an advantage to develop an application on the same platform on which it is to be deployed. "The parasitics in the model system are guaranteed to be directly comparable to the 'simulation,' " Minch said. By contrast, "In the process of developing an ASIC, one is never completely sure if the model is tuned up properly and what the parasitics actually will be when the chip is fabricated and packaged."

Educational opportunity
But the new technology may have the biggest impact in training a new generation of nonspecialist analog engineers. Twigg and Hasler have developed a system that allows students to go through a full analog design and test cycle in just a few weeks. Inspired by the educational boards and kits used by FPGA manufacturers, the system integrates various D/A and A/D converters and other interfaces to avoid the need for the testbench equipment normally used for analog test. Circuits are designed using XCircuit, an open-source program, and then can be tested using a Matlab graphical user interface. Hasler uses the system as part of his own Georgia Tech courses, as well as in off-site courses for professional engineers and others.

Olin's Minch is also on board. "I plan to use these devices to teach a first course in microelectronics and a subsequent pair of courses in mixed-signal integrated-circuit design," he said. "At present, in the microelectronics course, the students [can only] build and characterize simple analog cells. The most complex circuit that they can practically explore involves maybe a dozen transistors. Also, the dynamic performance of their circuits in the breadboard is nothing like it would be on an integrated circuit, because the parasitics are vastly larger."

Working with FPAAs, Minch said, "will enable the students to explore the behavior of more-complex circuits without having to make and debug an enormous breadboard. Also, the dynamic performance of the FPAA will be much more like that of an integrated circuit than the breadboard is."

In a follow-on mixed-signal IC design course, Minch said, "the FPAAs would provide an alternative to simulation for exploring ideas. Simulations often fail to converge, and accurate model parameter sets are very difficult to come by. The FPAAs will never fail to converge. And,while students may get their projects fabricated in a different technology, transistors in the FPAA will certainly not exhibit nonphysical behavior, as some Spice models may when their parameter sets are not tuned properly."

Fabrication is still useful, said Minch, "but the FPAAs would provide a nice platform to test ideas before they are committed to silicon."

Yale's Culurciello also plans to use the new boards in class. "They are a great tool to teach [with], because they allow you to design circuits will little effort and no wiring," he said.

Future challenges
Of course, a number of issues will have to be resolved before the technology can succeed. One problem, said Culurciello, is that the "added noise due to the increased wiring and switches can limit the performance of analog circuits and lower the signal-to-noise ratio." He said Hasler has found a methodology to reduce the interconnection noise, but obviously not to zero. Also, Culurciello said, to make the technology genuinely useful, the re- searchers need to be able manufacture larger arrays of devices and to optimize programming tools for speed.

Hasler is optimistic that the climate for the FPAA technology will only get better with time. "These configurable approaches are important as technology continues to scale down. Tooling, mask generation, etc. become very expensive as we go to smaller processes, and correspondingly, the design cost continues to grow," he said. "Therefore, the hope is to get as much out of every design as possible."

Hence the trend toward programmable and reconfigurable digital chips. "FPAAs should allow the ultralow-power analog signal processing to take advantage of these technology trends as well," Hasler said.

Sunny Bains (www.sunnybains.com/blog) is a scientist and journalist based in London.