Lowering Energy Consumption in Always-on IoT Designs

As a lead design engineer on the Tensilica R&D team, Dr. Avinash Lingamneni certainly understands the trials and tribulations of developing always-on applications for the Internet of Things (IoT), wearable, and wireless markets. His team develops Cadence’s Tensilica Fusion DSP. Over the summer, Avinash, who has a Ph.D. in electrical and computer engineering from Rice University, discussed energy-efficient signal processing for IoT applications at the Linley IoT Conference. Afterwards, we sat down to discuss IoT processors, cognitive layering, and Avinash's unique approach to lowering energy consumption in computers. Listen in.

In this always-on IoT world, what are some main obstacles to delivering a rich user experience without excessive power consumption?

During my Ph.D. studies, IoT and always on were just buzz words. Now, we see voice-controlled smartphones, a lot of mobile and health-related apps, even the concept of the smart home. One of the biggest stumbling blocks is that there is no one single processor system that fits all of this. An ideal processor needs to be flexible, scalable, and low energy. Given the challenges, it’s not feasible to have one processor to do all of these things. It would consume too much power; plus, the conventional OS is not designed for these types of applications.

How can engineers designing always-on applications overcome these obstacles?

At Cadence, we’re applying to our technologies a cognitive layering approach, where we partition a processing task into layers or states that can be addressed by an appropriate processing engine. This creates the illusion that the application is running all the time, but what’s happening is, each layer does just enough processing before seamlessly waking up the next higher energy-consuming layer, if it’s needed.

The lowest layer has to be the most energy efficient and it has to have basic functions, such as the ability to gather data from sensors as well as close integration with sensors. Now, in IoT, there is no unique sensor—there are sensors for motion detection, temperature, audio, facial recognition, and much more. So an important consideration is that an always-on system needs to integrate a wide variety of sensors.

After the data gathering step at the lowest level, the cognitive layering approach triggers the next level of processing: data analysis. Finally, after the data is analyzed, the system needs to efficiently communicate the results with something else—the cloud or another IoT device, for example.

We developed the Tensilica Fusion DSP with this cognitive layering approach—it is scalable, flexible, configurable, and meets the energy consumption requirements of a variety of IoT applications.

How is the Tensilica Fusion DSP suited to meeting the requirements of three always-alert functions: sensing, computation, and communications?

The Fusion DSP is built on a base Cadence Xtensa configurable processor, and customized for specific application domains, whether a customer needs low-end processing with high energy efficiency, something at the other end of the spectrum, or something in between.

If we design the most energy-efficient processor at the lowest level of the cognitive layering scheme, it needs to at least gather data from the sensor—that’s the sensing function. The processor then needs to process and analyze this data, which is the computation function. Finally, the processor needs to communicate with the cloud or other IoT devices. The Tensilica Fusion DSP does all of this.

Describe how a Tensilica Fusion DSP can be used in a particular IoT application, such as indoor inertial navigation tracking.

A subclass of IoT applications for the Fusion DSP is sensor fusion, where the DSP gathers data from various sensors, aggregates and analyzes the data, and comes up with insights. For indoor inertial navigation, when users move around in a particular room, for example, we typically don’t use GPS because it uses too much power or lacks a clear line-of-sight access to the satellites. Using the Tensilica Fusion DSP for sensor fusion, we can combine sensor data through algorithms such as Kalman filtering to track movement.

Tell us more about your role on the Tensilica R&D team.

I lead the DSP Engineering R&D effort on the Fusion DSP. We develop the microarchitecture, instruction set architecture (ISA) extensions, as well as the microarchitecture optimizations. I also contribute to microarchitecture optimizations on our advanced DSPs for baseband processing and vision applications.

What is it like for you to be able to work on technologies that, ultimately, result in products that people around the world will use?

I do get a sense of fulfillment when my work impacts everyday people, or when I can see that my efforts will appear in my own cellphone or a wearable device at some time. This kind of experience was lacking during my PhD work, where our research tended to be confined to research manuscripts and might not see the light of day. I made a conscious choice to go into more product-driven work, to do research as well as product development. That’s exciting!

What got you interested in engineering?

Growing up, parents either direct you to become a doctor or an engineer. I’ve always liked to go into the innards of how computers work, what kinds of things they are built on, how we can build more efficient computers.

I first got involved in related research projects while working on my bachelor’s degree, then started on a PhD track to develop better, more efficient computers. During my PhD studies, I’d advocated a unique philosophy of approximate computing where we relax the notion of having 100% correct computers in order to achieve better energy consumption. Users are limited by what they can see, hear, and feel, so, often, 99.9% accuracy is just fine and can result in a substantial energy savings.

What is the next big engineering problem that you’d like to help tackle?

Most of my work over the last six years has been to push the notion of “sufficient” processing at the benefit of getting performance beyond what is possible or lowering energy. People are recognizing that doing more than what is required is no longer the right approach to designing things. I’d like to drive this philosophy into all aspects of system design.

When you’re not working on Tensilica technologies, what are some of your favorite things to do outside of the office?

I have lots of friends in the Bay Area, so we enjoy hanging out and exploring the area through hikes and tasting the different cuisines. Whenever I get a chance, I also enjoy playing badminton, chess, and ping pong.

Christine Young