Almost any vision of the near future includes the proliferation of smart devices; smart sensors and processors will be enmeshed in our clothes, hidden in jewelry and even in body implants. Engineers are making great strides in creating the types of tiny, flexible and biocompatible circuits needed to realize that future. But the major elements that are still missing are the hardware and software that can operate on these small, ultra-low-power devices.
To work, these gadgets will have to sip tiny amounts of energy, in the range of tens of milliwatts. By comparison, a typical smartphone uses tens of watts, while desktop computers can use hundreds to run their processors.
“Even if it becomes possible to make the processor hardware ultra-low-power, these devices would be unusable if software developers are not able to write efficient programs that can run on these devices,” explains Joshua San Miguel, an assistant professor of electrical and computer engineering at the University of Wisconsin-Madison.
That’s why San Miguel will develop hardware and software tools for a new type of computing on ultra-low-power devices in a project funded by a five-year, $538,645 National Science Foundation CAREER award.
The project focuses on a type of computing called approximate or stochastic computing. Most computer processors read and write in binary, which performs calculations using just two-wide digits. Still, that type of processing requires thousands upon thousands of logic gates and a degree of processing power that likely can’t fit in a next-gen smart device.
But stochastic computing reads and writes in unary, in which data is encoded as probabilities instead of digital numbers. This computing style allows circuits that are orders of magnitude smaller than those needed for binary. There are tradeoffs; stochastic computing can return some inaccurate results. But in many cases, the accuracy of these processors is good enough for the tasks they will carry out.
“Actually, stochastic computing fits very well with the kinds of computations that you’d be doing on these smart devices, which is usually machine learning or analytics. Usually, you’re just collecting information from sensors, like temperature readings, which are already noisy,” says San Miguel. “Everything is already probabilistic.”
However, building processors that use stochastic computing is not easy; in general, there are no standard tool chains available for designing, simulating and comparing stochastic computing circuits. And there are no stochastic computing tools that allow interested programmers to write useful programs.
San Miguel hopes to change that by developing an open-source framework he’s calling UnarySim, which will allow programmers to design and evaluate stochastic computing systems. Eventually, he hopes his work will lead to the first general-purpose, programmable stochastic computing processor which can be used in all sorts of applications.
“We want to develop an actual device that’s stochastic and runs whatever task you throw at it, whether it be compression, data analytics, machine learning, et cetera,” says San Miguel. “It will be based on a design that anyone could build off of in the future. We think that’s the key to making stochastic computing a first-class citizen in the computing industry. Hopefully, super- low-power devices can become a realistic thing in the next decade or so.”
The CAREER award also has an education element that he hopes will expand knowledge of stochastic computing. San Miguel plans to expose undergraduates to stochastic computing by having them build low-power, stochastic computing versions of common apps, like video conferencing and augmented reality software. He and his students could then distribute those learning tools to rural schools where bandwidth and network restrictions make conventional versions difficult to use.
“Low power actually enables a lot of things,” says San Miguel. “We’re having undergrad students put into practice these concepts and then having them realize the impact they would have on a community by doing so.”