The University of Washington’s Processing Systems Lab is pushing the boundaries of energy-efficient compute. With ambitions to showcase its work on the most prominent real-world platforms, the Lab’s Visvesh Sathe explains why Arm Research is the perfect partner.
“In the Processing Systems Lab at the University of Washington, we are exploring a broad range of problems in energy-efficient computing. We want to enable efficiencies in everything from ultra-low power sensing – stuff that involves harvesting energy from ambient sources – all the way out to high-performance microprocessors for things that run in gigahertz.
Having done a lot of work with existing technologies, we are increasingly looking at what the future holds, and this leads us to specific problems of machine learning and neural interfaces. We work in power management for chips, and in baseband processing for communication systems, and one thing we value is that when working on a problem, we do not declare victory based on simulation results. Rather, we are interested in building out and demonstrating the solution in a hardware prototype that allows for a meaningful evaluation of the work. The IP that Arm provides us here is invaluable.
We recently worked on a new way of doing neural stimulation, for example. When we stimulate the brain, the devices have to be able to withstand a significant amount of voltage. We really wanted to stay true to what is required for a real-world implementation. It turned out we needed memory, but the kit we were given did not have any of the necessary IP. We were two-and-a-half months from tape-out, when we would be submitting the design for fabrication, and we had nothing. So I sent out a frantic email to Arm Research, asking them to give us everything they had in terms of memory and logic IP. We got it within a week, and within two we had started running our own synthesis flows. That really got us out of a tight spot.
The interactions we enjoy with Arm’s engineers are extremely helpful too. There is certainly a gap between industry and academia: these companies are the ones dealing with customers and the real-world challenges associated with tech deployment. Meanwhile, it can be very easy for academics to get siloed into a specific problem, to the point that the system-level considerations completely change on you over time, and you end up producing work that, while interesting, does not have the specific relevance you had hoped for.
So having a map that is informed by interactions with people out there deploying tech in the real world is invaluable, as is the feedback we get from our collaborations with Arm. Often they have influenced the direction the research took, or the assumptions that we made when we were working on solutions.
I recently had a call with the Arm Research team in Austin, for example, where we talked about problems in thermal management. It was very interesting to learn the various ways in which thermal constraints have presented themselves as a problem of interest for them. These real-world challenges may not show up in publications or news articles, so present themselves as unique learning opportunities as to what our next research project could be, before most people are even aware of it.
But our group right now is more limited by funds than it is by ideas. We have plenty of radically new concepts and projects we want to explore; the bigger challenge is securing funds for grad students to conduct their research.
Our collaboration with Arm has not taken the form of direct funding as yet, but it certainly has proven to be a key enabler in this area as well. What we work on tends to require hardware prototyping to prove the effectiveness of the solution or methodology we are proposing. If we want to build a processor chip to demonstrate an idea in clocking, for example, we need that access to IP. Being able to showcase our research on credible platforms has been an immense gift. It means we are able to be published in the best places – and that helps us secure that much-need funding.
Arm’s IP also allows us to build SOCs with less effort and in way less time, so our students can focus their time on the more novel aspects of the research, building solutions that are translatable into the real world – honing their skills in controls, signal processing, architecture or analog design, or power management. With Arm’s help our students can spend 80% of their time on the research problem, and only 20% of their time on building the prototype that demonstrates that tech. That is good for research, and extremely good for the student: they are not doing repetitive work; they are getting out their ideas there, in the places that count.
Another of the threads in our research relate to self-optimizing systems, building hardware that maximizes its own overall energy efficiency based on whatever load is provided, while also satisfying the system’s required performance constraints. There is been work in the past that is looked at minimizing energy, but that is had limitations: it was done on a much smaller test case, did not really minimize overall energy, and did not regulate guaranteed performance. Any real-world system requires you first to guarantee that performance, doing what you can squeeze as much energy efficiency out of the system based on whatever context it is operating in.
Our work involved a mix of power management, run-time control, and optimization. Usually people demonstrate these techniques on a synthetic load device. Thanks to Arm we were able to demonstrate our work on the low-power Arm Cortex-M0 processor. I remember telling an industry peer about what we had done, that we had demonstrated a functional microprocessor, a real-world system, that optimized its own energy. He was very pleasantly surprised.
But one thing I have learned is that I can have all the IP in the world, but unless there is an effective mechanism to train our students on using that IP, whether that is tutorials or reference designs they can look at, you cannot plug it in and play. It could take six or seven months to learn, which is a huge opportunity cost for a grad student. Arm has taken that on board. They heard me, they agree, and they are working on it.
I really do not understand how Arm has managed to be this supportive to non-paying customers. And it is worth noting that the company has gone beyond just supplying us with IP. It joined the centre for Neurotechnology, a joint enterprise in which the University of Washington is lead institution, and which has a require to advance neurotechnologies. By becoming a contributing member, Arm has shown that it truly takes a long view on how technology as a whole can be transformative.
It is all very exciting. Arm has been fantastic. I have had extremely positive experiences with these guys, and I am looking forward to continuing to work with them.
Visvesh Sathe is Associate Professor, Department of Electrical and Computer Engineering, at the University of Washington, Seattle.