Wireless Sensors: Self- Powered System Design for the Next Generation
The Internet of Things (IoT) was supposed to proliferate across as many as a trillion devices by 2015. Six years on, we are nowhere close to that number. Batteries are the problem. In this recording of the Rice Electrical and Computer Engineering Distinguished Speaker Series, Dr. Benton Calhoun offers wireless sensor technology as a solution. Always-on, system-on-chip Eversensors provide 24/7 continuous data in real-time. They never run out of power. Their self-sustaining data harvest and transmission will separate factories of the future from those of the past.
KION: Good afternoon everyone, it’s great for us to have professor Ben Calhoun from the University of Virginia to give us a talk. Professor Calhoun received his bachelor’s degree from University of Virginia and his Masters and PhD degree from MIT and since then he has been on faculty of University of Virginia where he is now a professor. His research emphasized energy efficient and substantial circuit design for self-powered, batteryless sensing systems; that’s also what he will share to us today. He has been an author of several books and two more than 200 peer-reviewed applications and 22 issued U.S. patents. He has been a Stat Director for an NSF and an engineering research center. He’s also the co-founder and co-CTO of Everactive, which is a startup that’s doing extremely well nowadays. So, welcome!
DR.BENTON CALHOUN: Great! Thank you, Kion. It’s my pleasure to be here today and to tell you a little bit about my research in self-power system design that’s aimed at next generation wireless sensors for the large scale internet of things. So this slide captures a little bit of the context of the internet of things the idea there is that we hope that devices will get disseminated into our world and into our environments so that those environments can really become smart. And we hear a lot about this in many different contexts for example, smart manufacturing, where we hope that sensors keep track of the goods that are being produced but also the equipment and the assets that are doing that production to make sure that they can be maintained in a predictive fashion so they can operate as efficiently as possible saving resources and so on. Smart government, you know that’s a tricky one, maybe that’s an oxymoron I’ll skip that one, nobody’s quite figured out how to make government smart. Yet smart health is certainly a hot topic with the idea being that ongoing continuous monitoring of our bodies can do a better job of producing positive outcomes by getting information to stakeholders like our physicians and ongoing and continuous fashion and this idea just perpetuates across all these different segments if we’re able to put electronic devices out into the world in a way that’s ubiquitous and unobtrusive that we can get access to information and connect that information to the people who need to use it so that we can make informed decisions and have our world operate in a way that is more smart and efficient than is the case today.
Is this really occurring though? Well unfortunately this large-scale internet of things is not coming to be as quickly as people were predicting. For example this slide shows a few predictions that were made. In 2012, IBM was predicting that by 2015 there would be a trillion IoT devices but that didn’t happen. In 2017 the prediction was down to 50 billion devices by 2020 but that didn’t happen, and then in 2018 the prediction had reduced further to 25 billion devices by 2025 so again pushing down the number of devices and pushing out the delivery date so the internet of things the large-scale internet of things seems like it’s still coming but progress toward that large vision has been slow. So, why is it slow? Well I have one idea this might not be the only idea but I think one of the big reasons is that batteries are incompatible with this huge scale internet of things and even some back of the envelope map can allow us to see the problem the batteries introduce. Suppose we waved the magic wand and produced a battery that had a lifetime of 10 years. We can’t really do that in most applications today, but if we were able to do it and deploy those 10-year batteries in a trillion sensors –the kind of numbers that were being predicted for the large-scale internet of things – you still need to change 274 million batteries in a day if the battery only has a three year lifetime then you’re changing almost a billion batteries every day. So even the back of the envelope map batteries are just not compatible with a huge scale internet of things they require intervention by people and that level of maintenance just isn’t going to work with a huge distribution of devices. You can see that on this plot as well which shows on the log scale the number of units sold on the y-axis against time over decades on the x-axis and you can really see the proliferation of devices as electronics have gone from laptops that were one per professional down to everybody having computers. Then of course in the post mobile world that we live in today where we each have in addition to a computer and a phone a number of other connected devices, but still on order maybe 10 devices per person in order to get to this trillion node internet of things vision we’re talking about a much larger scale, and like I said before there’s just no opportunity to get there with batteries so it really points to the inevitability of a batteryless solution.
I believe that’s true that it’s inevitable that eventually self-powered operation by batteryless wireless sensors is going to be how we get to this huge scale internet of things. The opportunity of this huge scale internet of things comes in the form of access to new kinds of data. This slide shows an estimate of the total amount of data in exabytes the next byte is a billion gigabytes and you can see that that in today’s world most of the incredibly large amount of data that we’re producing as as humanity comes in the form of unstructured content that’s produced by people, you know the kinds of things that we post online and photographs and so on. Here’s a sliver of information that’s generated by IoT type devices that are battery powered, but it’s not until we untether from the battery and get to the point where we can deploy these internet of things devices in truly huge scales that we can start to produce access to data on the same scale as this unstructured data that we have today. That will really get us to the point where our environments can start to become smart because we’ll have this access to these the untapped physical world, so I’m making this claim that the way to get to this large-scale internet of things is by getting rid of the batteries and using energy harvesting to have self-powered operation is that the answer to the power problem.
Self-Powered Operation as a Solution to the Battery Problem
Well it turns out I think it is, but we have to face some challenges first so interestingly enough with the exception of outdoor light where you can get lots of power per small area like a square centimeter, in most other harvesting modalities no matter what you pick you end up with a fairly small amount of power available, just tens of microwatts per square centimeter. That’s true of indoor solar light harvesting, that’s true of thermoelectric generators, that harvest from temperature differences that’s true of radio frequency harvesting from a nearby radio source and so on so we have to envision an active operating condition for the electronics that hopefully can take advantage of those tens of microwatts. But unfortunately, most off-the-shelf electronics in today’s world when they’re active are consuming milliwatts– which is a thousand times too high to really take advantage of harvested power unless we turn them off most of the time. And in fact that’s if you find a batteryless device in today’s world, usually that’s the solution it’s just off most of the time. So it’s not very capable because it’s not active continuously a more thorough solution that really solves the root of the problem is to lower the power consumption of the active operating electronic systems so that they only need to consume less than 10 microwatts or get down on the order of microwatts or nano watts if we are able to do that. Then we can match them with the amount of energy that we’re able to harvest from the environment and then we can have a truly self-powered system that is also continuously active.
So thinking about this kind of self-powered operation we should consider: what sorts of applications could make the most use of this self-powered device where we don’t need to use batteries? Well first of all it comes back to that back of the envelope map that I talked about a few slides ago. Really the applications that take advantage of self-powered operation are those that are very high volume, meaning thousands of sensors or tens of thousands or more of sensing points. You know if you only need to sense in a couple of places like your iwatch you’re willing to charge that battery every night, or I’m willing to charge my phone every night because it’s important and useful and it doesn’t bother me that much. But, if I had a thousand phones or 10,000 phones you can bet I’m not going to plug them all into chargers even once a year, it’s just too costly. Also of course self-powered lends itself to applications where you need to monitor over a long period of time and some of those application verticals, that I talked about on the very first slide, such as manufacturing or agriculture or smart infrastructure. All of course have needs where that those infrastructure and those assets are in place for many years or even decades and hints could make use of self-powered operations a subcategory of these high-volume sensing points are those devices that need to monitor in places that are hard to reach for one reason or another. Perhaps they’re very remote and inaccessible, or perhaps they’re dangerous because of hazardous materials or other unsafe conditions. But these are other situations where self-powered devices make a lot of sense because if you put a battery powered device in those kinds of contexts then a person has to go back into that dangerous situation to maintain them.
So summarizing all of this, the kinds of applications that I think about when I envision self-powered wireless sensors as the best answer are currently unmonitored or under-monitored environments places where it just doesn’t really work right now to have a good monitoring solution because you can’t run wire there to have a wired solution and it just doesn’t work to use batteries because for example, they’re too many devices. So the last thing is because of these capabilities you tend to have lots of work that you want to do. Meaning you want to have these devices operate in more of a continuous fashion if you’re willing to get by with sensing rarely then you might be able to use a battery because you can just have the device be off most of the time and allow the battery to achieve long life by doing less. Of course, that’s an essential characteristic of battery-powered operation that is often overlooked. The only way you get longer life from a battery is by doing less work so that you can stretch the lifetime out.
In contrast, if you’re harvesting power from the environment you can work continuously and get a lot more work done so long as the active power consumed is less than the power that you’re harvesting.. So in that context, I’m going to in the rest of this talk describe some of the basic operations of a self-powered system like a wireless sensor node and talk about some of the key trade-offs for getting these self-powered systems to operate. then I’ll give you a few examples of systems on chip that my research group has built for self-power systems. Then I’ll talk about a few of the special components for cell power systems predominantly to give you an idea of what’s possible in these different scenarios. I’ll briefly tell you about some commercialized self-power systems that my startup Everactive has in the market today and then concludes the talk.
Basics of Self-Powered Operation
So first let’s talk about the basics of self-powered operation. I like to use an analogy where energy that we’re harvesting from the environment is analogous to water that you’re flowing into a tank. So the harvested power is like the water that you’re getting from the environment; maybe it’s rain water, maybe it’s from a spring and then you’re storing that water.for example, in the tank. In the electrical context this would be like a capacitor or super capacitor and then of course when you use that energy you’re drawing water out of the tank. This analogy makes it pretty obvious to see some basic tenets of self-powered operation. If you’re harvesting more power than you’re consuming, then of course you’re going to fill up the tank and store energy in your energy storage medium in that context. When you’re harvesting more than you’re using, you’d like to optimize for power because if you can keep the load power lower than the harvested power then you can be in this energy rich condition for a longer period of time and more of the fraction of the node’s lifetime. If on the other hand, the amount of power that you’re consuming is larger than the load then the harvested power is less than what you’re doing then you’re gonna use up the energy that you have stored in the tank and eventually the node’s going to run out of energy and die unless you change something.
So one thing you can change is to work less to try to reduce the load power. But in this context, you’re using the stored energy so you really have to optimize the system for minimizing energy and that’s subtly but importantly different from optimizing for power. So the last thing to recognize in the self-powered context is that the source of the harvest, the power, is not something that we can control. It comes from the environment and it tends to be erratic. There’s some sources of power like sunlight that are periodic or semi-periodic but a lot of the time the amount of power that we can harvest is something we can’t control and it changes in ways that are hard to predict. Which means that we have to have a system that’s self-powered that operates in a way that is smart. It needs to be aware of how much power is available it needs to be aware of how much energy is stored and it needs to be aware of how much power it’s consuming so that it can make the best decisions about how to operate in its self-powered context.
So this slide shows a rough hierarchy of a self-powered system. Both the sensor node down near the bottom of the slide which is going to be the predominant focus of my comments today, but it’s important to remember that that node is a part of a larger system. It exists in the environment with multiple other nodes. Because of the nature of self-power systems there are probably thousands of other self-powered nodes nearby. And therefore, it needs to communicate with those nodes and with the back end of the system through a network and of course back at the back end perhaps that’s in the cloud or on a gateway. So on there is an additional application function and the rest of the larger system. I won’t focus on those aspects of the cell power system. I’ll predominantly be talking about a self-powered sensor node today but it is important to remember that they’re there because they do influence what happens with the self-powered node design.
Now if we look at the blocks involved in the self-powered node a few of these key blocks I’ll dig into in more detail during the rest of the talk. Of course this node has to harvest energy; that harvesting happens through an energy harvesting transducer that converts some energy in the environment into electrical energy. For example, a photovoltaic cell converts light into electrical energy, then of course there’s a harvesting circuit that pulls that electrical energy out of the transducer and conditions it, perhaps dumping it into an energy storage node or regulating it for use by the rest of the electronics in the system. Those circuits tend to be on ship, the energy storage and the harvesting transducer perhaps could be on ship, but they tend to be off-chip discrete components. Then there are the load parts of the system: there are some circuits that are always on, and that’s an important part of the cell power system because they can really limit the bottom end of the amount of power consumption that the system consumes. And then what we do with that system can vary, but typically, there will be some sort of sensing elements with the associated sensing circuits.
The digital parts of the system that compute and store information from that digital signal that comes from the sensors. Then communication, perhaps in the form of radio communication to talk to other parts of the system through the network. So a slightly different picture of this self-power sensor node is shown at the top here and I’ll use this this picture repeatedly through the rest of the talk to illustrate what part of the system I’m talking about you can see on the left there’s the environment which interfaces with the harvesting materials, then the harvesting circuits, the energy storage, the energy that’s stored on that energy storage node, gets regulated down to have clean voltage supplies for the load circuits, which of course have underlying them the power management and the always-on part. Then for the data path there’s memory processing, sensing and communication.
Designing System-on-Chips
So the next section of the talk will dig into what a system-on-chip for a self-powered system might look like. And then give several examples from the work that that my research group has done at the University of Virginia that walk through a couple of generations of self-powered systems on chip or self-powered systems. That walks down the power consumption from about 20 microwatts to half a microwatt.
So first let’s think about all of the different pieces that would go into a system on chip for self power system operation. So this is a generic block diagram of a system on chip that might be used in this context. I’ll walk through just a few key pieces of it. Of course the sensors interface with the world to get information from the world that you’re interested in detecting. Those get converted through an analog front end and an analog to digital converter to bring that information into the digital domain, where they’re typically processed using a general purpose processor or microcontroller of some sort.
Some of these microcontrollers act in an always-on fashion and others are duty cycles meaning. they’re off a lot of the time but then turned on to do work for brief periods of time. Frequently, the always-on function of the processing elements is to perform node management; including perhaps power management of the energy harvesting and the used power by the load. Sometimes there will be digital accelerators that are specialized to do processing that occurs repeatedly and that is specific to the function of the particular device. Of course the underlying energy harvesting and power management circuits need to perform the functions that I already talked about: of bringing in and conditioning the energy that we’re getting from the environment, adjusting to the different needs of the load circuits, and then trying to optimize the flow of that energy through the system.
On the communication side, we’ve found that wake up radios are a critical part of a system on chip for self-power systems. I’ll dig into wake up radios briefly later in the talk. But they provide these systems the ability to remain in low latency connection to the network but at power levels that are consistent with with the kinds of budgets that you can harvest from the environment.
Of course, you probably also will need some sort of communication transceiver that does both transmit and reception that does higher bandwidth higher data rate communication when necessary. This communication radio might be standard compliant. And it probably is one of the dominant peak power hogs of the system; so you want to optimize the system to use this radio as infrequently as possible to minimize the total energy that it contributes to the power budget. And then there’s some always-on functions like clocking to generate the bias voltages and currents and references and clock sources that the system needs to function as a system.
So I took the time to walk through these different pieces just to give a sense that a system on chip for self-powered operation is actually by itself a fairly complicated system that has a lot of interdependent components.
This slide speaks to what’s available with off-the-shelf components in electronics in today’s world. You’ll see that there’s some things here that are green, meaning some of the off-the-shelf components that you can buy today are quite appropriate for use in self-power systems and work really well. For example, the power consumption in many different kinds of sensors that are used in cell phones has been reduced dramatically down to the microwatt level in many cases, simply because of the positive market pressure of having really high volume in cell phones. It really pushed the power down. Likewise, there’s some microcontrollers that are really good at duty cycled applications; meaning they have low off power and very low standby power. They can turn on rapidly and do a lot of active computation very quickly, but when they’re on they tend to be high power relative to a self-powering context, meaning milliwatts.
You can see that some of the other available off-the-shelf components are not as amenable to operating in a self-powered context. For example, radios tend to be in the many milliwatts of power consumption when they’re active if you’re getting them off the shelf.So there’s a lot of room for improvement. If you remember the amount of energy that we can harvest from a store from a harvesting environment, milliwatts is a hundred x to a thousand x too high power for what we’re looking for for continuous operation.
I am not going to walk through this slide it’s very busy it has lots of different points each of which is related to some published work that we’ve had but if you’re looking at this on video you can pause the video and take a look and if you’re interested in one of these particular components feel free to reach out and we’d love to share more information about it but I wanted to show it just to give the sense that that my work my work in my group has targeted all of these different components of required for self-powered operation and been able to demonstrate deep sub-micron active operation of these components and of course there are many other people contributing to these kinds of components as well.
So in the research community, there’s a lot of work that has advanced the power consumption well below the microwatt level that we need to make self-powered systems realizable. Let me now walk through three examples starting with the first integrated system on chip that we built back in 2012.
First System-on-Chip designs
For Body Nodes
This was for wearable applications and it was an on-body, batteryless, body heat powered chip, that consumed an active power of 19 microwatts while monitoring continuous electrocardiogram signals. So that’s the electrical waveform of your heart and you can see a measured example of the ecg signal from this chip here at the bottom. It detected atrial fibrillation —which is a cardiac arrhythmia — using a digital accelerator on the chip that implemented an algorithm produced by a cardiologist in the afib center at the University of Virginia hospital. And within just the first 10 beats or so of afib it could detect that signal or that condition and then send a radio update to communicate that that condition was occurring. All with an average power of 19 microwatts doing that continuous operation.
You can see at this measured waveform how in this particular chip we used an RF pulse to wake up the chip when it was in a zero power condition. That turned on the boost converter, which then started to harvest from a thermoelectric generator, which converted the thermal difference between the skin and the air into a usable voltage that we stored on a capacitor.
For Body Nodes and IoT
The second example of the system on chip came a few years after that and was an evolution of that first design. We designed this system on chip to be quite a bit more flexible. It still can target on body applications like EKG, it actually has the same analog front end as the previous system on chip. But it was designed to be far more flexible and support a wider range of different application spaces for the broader internet of things (IoT). So it includes a number of different accelerators and different sensing interfaces and so on. And we were able to show a total active power of the system on ship of just about six and a half microwatts, including continuous 187 kilobit per second transmission by an ultra wide band integrated transmitter.
So this design shows that particular choices in the radio can achieve very low power consumption and modest data rates like 200 kilobits per second. One advantage of this ultra wideband transmitter is that it could support high data rate of up to a megabit per second and only push the power up to about 30 microwatts, still in the range that can be supported by self-powered operation in a lot of environments.
Full System-on-Chip
The final example that I’ll show of a full system on chip is this design where we actually decided to partition the system and put the radio components on a separate chip in the package and the non-volatile memory component on a third chip in the package. In that case, because it was a special technology that supported ferroelectric ram. But in this design the core system on chip consumed half a microwatt so 500 nanowatts and the non-volatile memory which was custom was only four microwatts of active power and this whole system could be put together with an active power on average of less than a microwatt. So again just walking down the power of these components makes a big difference but it is also critical in order to design the system from a system viewpoint starting with what the application’s needs are and propagating those needs down to the component design.
Key Components and Trade-offs of Self-Powered Systems
Alright so now I want to walk through a few of the key components and give you a sense of some of the key trade-offs and some of the options for designing those components that are essential to allowing self-powered systems to operate. I’ll spend most of my time talking about the harvesting circuits and the dc to dc regulation circuits. But i’ll tell you a little bit about some of the load circuits as well just so you get a sense of what these things look like in cell power systems.
So again starting with the energy harvesting and power management, I’ll talk about some of the key trade-offs and how these circuits behave. So of course it’s important to think about the ambient environment and to select the energy transducers that are going to take the type of energy that’s available in that environment to provide you with harvested power.
Photovoltaics of course, work well in conditions where there’s light indoor or outdoor. And the choice of the photovoltaic material will vary depending on whether you’re going to be operating more in indoor conditions or outdoor conditions.
Thermoelectric generators (TEG) convert a thermal difference into a dc voltage, and so those can be used in context where you have a source that is going to be hotter or colder than the ambient condition.
And there are many kinds of kinetic harvesters such as piezoelectric materials, or micro electromechanical systems, that can harvest kinetic motion and convert that ac waveform into after a rectifier a dc harvestable waveform. And then of course you can use coils or antennas to harvest different electromagnetic energy.
Energy Harvesting Circuits
The energy harvesting circuits take the non-conditioned waveform from the energy transducer and turn it into some usable energy that can be pushed into for example, a capacitor for storage. The most common energy harvesting circuits are a boost converter, which takes— think of it as in the water example — water out of a non-usable source like a puddle and bucket and brigades it up to fill up a tank for example. that would be one kind of a boost converter.
Rectifiers take oscillating inputs like ac inputs from a piezoelectric harvester, for example and regulate them to produce a dc input that you can then feed into a boost converter and use that for energy harvesting. It’s important with energy harvesting circuits to manage features like cold start which means how the system will start up when it has absolutely zero power to begin with. And maximum power point tracking which refers to making sure that the harvested circuit interfaces with the harvesting transducer in such a way that it pulls the most power out of that system as possible. Another way to think about this is impedance matching with the energy transducer.
The Water Example
And then of course the power management unit takes the stored energy and produces nice clean regulated supplies for the rest of the circuits. Again, in the water analogy if the harvesting source is something like a puddle, or a spring, or rainwater that’s intermittent, the boost converter takes that energy and stores it in a tank and then the regulators the power management unit produces a nice clean regulated ready-to-use source of power for the rest of the system. Which might be analogous to the water that comes out of the faucet when you turn it on.
Three Common Topologies
So just for your awareness i’ll talk about three very common power converter topologies the circuit diagrams here are shown for dc to dc converters but but similar topologies can be used for harvesters. An inductor-based voltage regulator uses an inductor which in the water analogy is a bit like a water wheel. If you want to make water run uphill one way to do it is to get a water wheel spinning by running water downhill and then flip the direction that’s coming out of the water wheel and that water still that water wheel still has a lot of momentum and it’ll push the water uphill for a bit. That’s analogous in a way to how an inductor works to in a boost converter context.
A switch capacitor voltage regulator uses a capacitor and a bunch of switches much like a bucket brigade. again back to my analogy, it’s a bit like taking small buckets of water out of a puddle and dumping them up into a bigger tank and gradually filling the tank. you can use switch capacitors to both regulate in the boost converting fashion or to regulate a nice clean output voltage.
And then a linear voltage regulator which i’ll give a few examples of later is really just a down converter and it’s analogous to usingyou know a faucet to regulate the flow of current through a resistor through a switch to get a nice clean supply at the load.
Key Parameters for Designing Good Regulators
So for the circuit designers in the crowd, I’ll mention one of the key parameters that’s really essential to designing good regulators. Whether they’re boost converters or down converters which is the power efficiency it’s measured as the power out divided by the power in and of course in an ideal regulator all of the power that goes into the regulator also comes out so you’d have an efficiency of 100 in reality there are some losses but you want to minimize those losses to keep the efficiency as high as possible the quiescent power is the power that’s consumed by the circuit when nothing’s really going on it’s a quiet environment there’s no real input it’s just the power consumed by the circuit while it’s sitting there so in a circuit this includes leakage current and bias current and things like that the switching loss is due to switching in the circuit and the conduction loss is due to current flowing through switches or through resistances
One key parameter in the design of sub microwave deep sub microwatt boost converters and dcv regulators is quiescent current. And I want to say a few things about that and then look at the trends of quiescent current over time. So when we’re building energy harvesters we’re interested in the worst case environmental conditions. And those conditions are of course going to produce outputs where we have less power available and the efficiency during those conditions is going to occur or is going to be limited at the low end of the power efficiency curve. So this plot shows a typical power efficiency curve against output power or some sort of a regulator and most historically most people are trying to have a maximum efficiency at a high output power because they’re interested in the system where they’re doing a lot of work. but in a self-powered system you’re actually trying to push the active power as low as possible even into the sub micro range. For traditional harvesters and regulators, the efficiency tends to be quite poor there. So this introduces a challenge for circuit designers to increase the efficiency down at the low end of the power efficiency curve and one key parameter for achieving that is to reduce the quiescent power.
So I already described how in typical conditions you’re going to harvest about 10 microwatts per square centimeter in poor conditions that amount of power can reduce even further to deep sub-microwatt. So to extend the operating range and the reliability of self-powered systems it’s important to reduce the quiescent power of regulatorsand and boost converters down or to increase the efficiency at load powers down in the sub micro up range. And if we look at a plot across time on the x-axis of the log of power consumption on the y-axis the very top of this plot 10 to the three nanowatts that’s a microwatt.
At the beginning of this talk I was describing the need to get below microwatts but you can see that in the last decade or so the power consumption for different components has decreased down to the nanowatt level and even below into the picowatt range. and for full systems this is also the case. you can see systems that are occurring have active powers of around a microwatt or half a microwatt like the one that I showed you earlier.
In order to provide energy efficiently for those kinds of components we need to also reduce the quiescent power of the regulators that are supplying energy to those components. And you can see that there is a trend in the literature over the same time period of the quiescent power of boost converters and regulators going down towards the nanowatt level or even below. And this is going to be important as we build self-powered systems for these hostile environments. This slide just shows a couple of examples of that on the left is an analog low dropout regulator that’s sub nano ampso less than a thousandth of a microamp that uses an analog topology which is zero ripple, although it can be slow in terms of its transient response. On the right is a different design that uses the hybrid approach in the digital domain to produce very rapid responses with still low ripple by using this hybrid design topology and again the quiescent current is less than the single nano amp.
This slide just shows some examples of those sub nano amp level quiescent power regulators that I was describing on the previous slide. So in addition to reducing just the quiescent power, it’s important in the sub microamp energy harvesting and power management units to consider some of the other trade-offs.
Trade-Offs of Design
For example, the trade-off of the transient response and the dynamic range and the stability of the design in terms of its performance across process variation temperature fluctuation. Which of course is going to be an issue in internet of things applications that are deployed across a rich set of environments and different voltages. So the energy harvesting power management architecture, if you use an inductor-based design, can be split into two stages where there’s a separate energy harvesting stage feeding an energy store like a capacitor and then a second regulator stage that uses a second inductor for example to produce the regulated output voltages so that would be a two-stage design.
A disadvantage of this is that it needs more inductors which are typically off-chip components, but even if they’re on chip components you’re cascading two stages. So the efficiency is the product of the efficiencies of those two stages. So there’s some efficiency loss there; you can do a design that uses a single inductor and does one stage of conversion from the energy harvester to either the storage node or the output depending on the conditions that can produce a higher conversion efficiency but it also leads to higher complexity and can be a more challenging design to get right?
I want to go quickly through a couple of examples just to show that there are solutions to these different options. This is a single stage design that uses one inductor for both harvesting from inputs as low as 10 millivolts. And using different types of sources at the harvesting end but that uses a single inductor to produce energy that goes either to a storage node or to one of several different regulated output voltages. Using that single inductor with multiple outputs the end-to-end efficiency all the way from the harvesting input to the output can be as high as 75.
Here’s a different example of an integrated energy harvesting power management unit. This is actually the design that is integrated on that 500 nanowatt system on chip that I showed earlier you can see that it’s a multi-stage design the total quiescent current is down around one nanowatts quiescent power is around a nanowatt and as low as 390 picowatts it consists of the first stage that has a couple of different options a switch capacitor harvester or an inductor based harvester the choice of which depends on the operating environment and then it also has several options on the regulation side to produce three different power voltages that include integrated switches for deciding when to deliver those voltages to off chip components and you can see in this measurement as the stored energy is increasing these regulated rails come up and and once the rails are all regulated through the dc-dc part of the system the system power on reset is triggered
Load Circuits
So I want to switch gears and talk briefly about the load circuit. So this is the part of the system on chip circuits that are doing the active work that’s required for the application. So these of course require careful co-design with the energy harvesting and power management unit to make sure that the two go together well, and one of the key pieces of these systems is the always-on piece. Other pieces that are not required to be active all the time should of course be duty-cycled, which means turning them off when you don’t need them so that they can reduce their average power as much as possible. I mentioned earlier the idea of using accelerators, which are specialized hardware for doing some sort of common computation efficiently, and it’s certainly not a new idea that customization is efficient. But this slide puts some numbers on that so you can be aware of the power of using a custom block, an asic-like block like a digital accelerator.
These comparisons are between an already power optimized sub-threshold design for a microcontroller. Compared with a custom digital accelerator that’s specific to that particular function, you can see that specialization gives you an extra two to three orders of magnitude of improvement in energy efficiency. So if you have a specialized design that you know is going to be doing something very frequently you should consider using a digital accelerator for that class of function because you’ll get two or three orders of magnitude of energy efficiency from using that accelerator. Of course, the cost is in flexibility because by definition customizing something for a particular function reduces its flexibility to do other functions. Here’s an example of a work we just published that’s a flexible analog front end designed with a current mode channel, a voltage mode channel, and a resistance mode channel. It’s designed for on-body applications to measure concurrently the electrocardiogram, your heart rate signal, your pulse oximetry signal so ppg signal which which shows you both pulse and heart rate, but allows you with the EKG to calculate the transition time of the pulse and get a a measure of blood pressure. And then at the same time, it can measure an environmental parameter like the concentration of ozone which is known to be an inflammatory material that can cause respiratory problems and exacerbate asthma. And things like that, this design just shows that it’s possible for three different types of channels concurrently operating for PPG, EKG and ozone measurement all the entire front end doing all those those three things at the same time is only 785 nanowatts, so less than a microwatt. So some advances in these mixed signal designs have made it possible to do really sophisticated continuous sensing at the sub microwatt level which is consistent with self-powered operation.
Always-On Circuits
I do want to say a few things about always-on circuits and I’ll start with wake-up radios. I mentioned this before, we’ve found that wake-up radios are the best way to keep these self-powered systems in low latency continuous communication with a network. The traditional way of staying in communication with the network — and this is the way your cell phone works and wi-fi works and bluetooth works —is to turn on a high-powered receiver and then turn it off again and leave it off for a long time and then turn it on again. And the process of making sure that when you turn it on, the thing that’s talking to you is ready it’s called synchronization and it actually consumes a lot of power to remain synchronized so that you can do synchronization well.
Well the wake up receiver gets rid of that problem by just leaving a receiver on all the time. Unfortunately a few years ago wake up receivers had pretty poor performance but there was a DARPA program called the Enzir program that was trying to push on this and improve the performance of wake up receivers. This slide just shows in summary the sensitivity of wake up receivers in decibels dBm on the x-axis against the log of power consumption on the y-axis. When the program started the best designs were on order 100 nanowatts and minus 40 dbm and over the course of the program they improved all the way from minus 60 dbm to minus 80 dbm to better than minus 100 dbm. We were able to show results that were over 10 000 times lower power than at the comparable sensitivity and a 10,000 time improvement in sensitivity. So now wake up radios can be used at the nanowatt level and can have sensitivities that allow them to get ranges in the kilometers or even better. So these are new categories of radios that are really important for self-power systems.
Clocks
I mentioned clocks as one of the always-on components that are necessary for operation of a self-powered system. I’m not going to go into the details on this slide except to show that the state of the art has really improved and provides a good trade-off space of different power levels down into the picowatt kind of range across frequency and maintaining different stability. And really the trick with clock generation is lowering the power while also keeping the accuracy of the clock even as things like temperature and voltage vary. Just one example design published just this month of a 10 nanowatt but very stable design that set a new record in terms of the clock efficiency in energy per cycle.
Sub-Threshold Operation
So for digital operations like microcontrollers, I do want to briefly mention this term sub threshold operation. You heard it in my bio, this is something I’ve looked a lot at. It refers to the operation of a transistor in the region where the gate voltage is less than the threshold voltage, called the sub-threshold region. Traditionally this would be considered as the transistor being off, but for the last couple decades people have realized that you can actually operate the transistors in that region and this slide shows a trade-off that would suggest why you would do that. The delay goes up dramatically when you go into the sub-threshold region so the transistors get quite a bit slower, exponentially slower, but the energy reduces quadratically as you reduce the voltage. This is the reason, the whole reason, people would try to operate digital surfaces of thresholds is to take advantage of this reduction in energy which can give you substantially less energy per operation for your digital function.
There’s been a more recent advance that additionally inserts leakage reduction circuits right into each gate which incurs a higher area overhead and further slows the circuits. But we’ve been looking at how to combine these extremely low leakage but cumbersome and very slow dynamic leakage suppression gates with traditional CMOs to produce a scalable design. We were able to successfully span the difference between these DLS processors and normal CMO processors and demonstrate a risk. 5 Microcontroller Risk Vibe is an open source instruction set architecture that can operate with power as low as less than a nanowatt, 800 picowatts, but then scale its performance up to the tens or hundreds of kilohertz at power consumption of around 100 nanowatts. This kind of a design is both very flexible and tunable post silicon, but also quite appropriate for a system-on-chip for self-power systems.
Self-Powered Systems in a Commercial Context
So I’ll take a little turn here and spend a few minutes talking about how self-powered systems might look in a commercial context. I’ll use the example of Everactive which is a startup that spun out of my lab at the University of Virginia and my co-founder Dave Wentzloff ‘s lab at the University of Michigan. So what Everactive has realized about cell power systems is really the same thing I described in the opening part of this talk, which is that the real value is in monitoring things that are currently unmonitored and what customers would like to see is access to those data streams. And in some sense they don’t even care that there are no batteries, they just don’t want to have to do maintenance on these devices. They have thousands and thousands of assets that they want to know what’s happening, and they basically need to see the data. So what Everactive has done to try to address that problem is: to use low power circuits to build batteryless, self-powered systems for the wireless sensors, but then connect those wireless sensors in very large numbers in a network that supports high densities and large numbers of devices and flow the data from those devices all the way back into a cloud platform where that new data stream can be made available to the people who want to see it.
And one example application that we have a product in the market to serve is for machine health monitoring specifically talking about machines like rotating machinery and electric motors. There are over 300 million electric motors in the world. Those electric motors when they operate, even though they can be expensive, most of the cost of their ownership is the electricity that they use during their lifetime. In fact, amazingly, electric motors consume almost half of the global electricity use. These things are used all over the place and motors, pumps, fans, gearboxes, all kinds of applications and they are a significant source of energy consumption. When they get misaligned and start to run in a more ragged fashion, they start to vibrate, they use more energy, they become inefficient, it costs lots of money and it can produce costly downtime in whatever process they’re supporting. So the ability to monitor these very large numbers of rotating assets is important in order to monitor them properly and provide the data back to those people who would take action like I mentioned. We need to have a sensing device that is part of a network of very large numbers of devices that are flowing that data to the end solution and that informs the requirements for the chips and the cell power system itself in the node
Everactive has tried to optimize the network for this to take advantage of wake up radios and to provide a good balance between the ability to have these devices on all the time. Which means they’re producing lots and lots and lots of data and many many devices up to a thousand per gateway in the small area, while still allowing each of those devices to have low latency access to the network. So over a long range 250 meters non-line of sight in an industrial environment, we’re able to connect a thousand devices per gateway with a pretty good data rate. So there are a lot of other protocols out there that have strengths for different applications but none of them actually meet the application needs of the large-scale internet of things. Some things like Zigbee and bluetooth can support really high data rates but they don’t have a very good range other protocols like Lorawan or Sigfox or WirelessHart. You know it might be able to supply better range, but they’re very low data rate and much higher peak power and really targeted for infrequent monitoring access instead of continuous operation. They’re really checking in just a few times a day so this of course influences what kind of chip you build.
I’ll just reference a paper at this year’s international solid state circus conference from Everactive about this commercial system on chip for self-powered system design. You can see from this picture on the left it uses a two-stage energy harvesting and then power management design with a single inductor multiple output regulator for producing multiple different voltage rails. You can see that the power consumption is all about the total system power, including all the sensors and everything going on in the application. Not so important, what just a piece of it is: the design does leverage accelerators for some of the commonly used functions in the machine health monitoring application. And it’s very flexible across different energy harvesting modalities and different features that make it amenable to actually being deployed in the real world. On the radio side, the wake up receiver allows us to integrate it in this high density network like I described, while also dealing with some things to make sure that it’s secure and robust against different kinds of cyber attacks.
So just to conclude, cell power systems are possible and they’re made possible by bringing the component power and the full system on chip power down to the sub microwatt level. And recent work by my group and many others has pushed that even further into the nanowatt scale or even the picowatt scale. The next challenges I think come from taking these components and viewing how they go together into complete self-power systems and that really requires that we treat self-power systems as a new hardware platform, think about how to model and optimize those self-power systems, and then how to deploy industrialized versions of them.
So I hope that this was an interesting introduction and got you thinking about self-power systems. And I hope that you are a part of what I believe is the next computing revolution which is the design of batteryless, energy harvesting cell power systems.
Question and Answer
Dr. Calhoun: So with that, I’ll stop and be happy to take any questions that you have. Thank you, so please use the Q&A hand on YouTube please.
Question: Hi Benton, thanks a lot for the nice thought. I’m serving as an assistant professor in the department. The product is really informative and the work is impressive. Our one question I have is in the pipeline shown earlier on the second vlog the loading circuit. So, what does the pie chart open look like? I think you give some information in this line 30, but for the IoT device of your particular interest could you show a particular high level wheel? What does the pie chart look like in terms of the memory processing and communications?
Answer: Yes, okay I think I understood your question. You’re asking about the pie chart, or the breakdown of power among different components is that right?
Asker #1: Yes
Dr. Calhoun: Yeah, it’s a great question and there’s not a simple answer. So the pie chart can be very application dependent with a couple of them. The highest peak power components are certainly the radio, whenever it’s on it you know can be a milliwatt level, or hundreds of microwatts kind of level device. The transmitter for example, and then frequently the sensor interface, can also be very high peak power. The pie chart of the average power across the system tends to be influenced by the application very heavily. For example, if you’re trying to stream a lot of data it’s going to be almost all transmitters. If you’re trying to do a lot of node edge computation like inference, it might skew more toward digital power. If you’re really only checking a few sensor elements and thresholding them and looking for some sort of rare but simple event, then the power can be dominated by just the always-on stuff like leakage in the clock. So it’s a very application independent pie chart.
Asker #1: Yeah, so the breakdown makes sense so in the in representative IoT application, or the one you are younger or more interesting? Could you provide more details for example in the surveillance IoT devices?
Dr. Calhoun: Sure so you’re asking about how the breakdown looks in some of the applications that I’ve worked on?
Asker #1: Yeah.
Dr. Calhoun: So if you know what your application is and how you’re trying to address the application needs, then the best pie chart breakdown will evenly allocate power across some of the different components like always on sensing processing and communication. So when we’ve done our job well in a given application, you’ll find that the contribution to those areas always on sensing processing and communication is pretty evenly distributed.
Asker #1: Thank you!
Question: I have a question. So I noticed that you mentioned that the SOC is operating at the sub threshold voltage like 0.5 volts. What limits you from going lower? I’ve seen some charts showing that you can go down to like 0.2, and the noise floor is low. Can you go lower than 0.2?
Answer: Yes, that’s a great question. So the first part of the answer is the voltage at the limiting voltage is going to be dependent on the actual process technology. So some of the papers you see where they operate at 0.2 volts, they might be newer smaller feature technologies where the threshold voltage is only 0.25, for example. So we tend to operate in older technologies with higher threshold voltages, partly to take advantage of the lower leakage that comes from using those older technologies. But I think part of your question was also about what limits at the low end of voltage. And really, the ultimate limit comes from when digital circuits can no longer retain their buy stability and and you can’t have the difference between a one or a zero because you’re basically running out of gain. And yes, people have demonstrated digital operation down below 100 millivolts, you know, in the sort of 60 millivolt level. Ultimately, you’re limited by the sub threshold slope of the transistors in terms of the lowest voltage, but there’s a practical consideration you know at some point. If you keep going lower and lower and lower you actually are not doing yourself any good. If I can find that slide again, the sub threshold curve if you notice, is the energy as a function of voltage. The energy reduces as a function of voltage but then it starts going up again. So if you go too low, you’re actually spending most of your time sitting around leaking and doing nothing because the circuits are so slow and you’re not actually going to win by being at that lower voltage. So, where you actually want to work is at this optimum energy point if your circuit is predominantly digital, and that’s one reason that you don’t see people trying to operate at much lower voltages.
Asker #1: Interesting, thank you.
Question: Yeah, so here there’s another question from Dr.Cavanaugh, he asked: have you considered a biomedical implant? Which would be more complicated if you have these scenarios?
Dr. Calhoun: Yeah, that’s a great question. I have not personally done any work on implants, partly just because of the regulatory complexity of putting something inside of a person. But I think there are certainly applications where you could think about using these same sorts of circuit techniques to work on implants. The low power consumption would certainly benefit a battery-powered implant, but there are technologies for doing batteryless implants. A lot of those tend to be remotely powered rather than energy harvesting, meaning you conductively or inductively coupled power through the skin, for example. But there are some examples of people investigating technologies for actually harvesting inside the body. One really interesting work from my advisor’s group at MIT actually uses a voltage potential that occurs naturally in the ear to harvest from. So there are some interesting areas there, but I’ve not personally worked in that area.
Asker #2: Thanks.
Question: Is there any other question? Okay so if not, I will have a last question for you: so in a highly duty cycle system the average power can be very low and match your harvesting source. But, you also need to support the instantaneous high power. Is the solution just having a large enough capacitor or are there more considerations? Because if you have a smaller capacitor, then the voltage drop on the capacitor will be large or if you charge it to a very high voltage when you’re doing the regulation there will be more loss to generate a regulated supply voltage. So what’s the consideration here?
Dr. Calhoun: Yeah, well it’s a complicated question because there are a lot of variables involved. As you suggest, I like to work on designs that are more continuously active and not duty cycled so much. But you’re right that you can solve certain kinds of problems where the need for doing work is infrequent. You can just use a duty cycle and have a very high peak power component that turns on rarely, does its work, and then turns back off. And you are correct at pointing out some of the trade-offs there. So you’ll need to store enough energy for allowing that particular work event to occur. And there’s a trade-off for how to do that with the capacitor, and what kind of capacitor to pick, and how much capacitance to pick and frankly you know maybe a battery makes sense if the power is high enough, although batteries are not particularly good at delivering a lot of current. So the best solution might be a harvester that charges a rechargeable battery. The rechargeable battery charges a really big capacitor and the capacitor gives the peak current deep draw of current to the high load high peak power component, but a lot of the right answer depends on the details.
Asker #3: Yeah,okay. Thank you!
Presenter: So that concludes today’s talk. Thank you for attending and thanks for the fantastic talk!
Dr. Calhoun: Thank you so much for having me, and hope everybody stays safe. Thank you!
For more information on Everactive’s self-powered wireless sensors, click here.