Experts at the table, part 2: How much more power can we save in designs, and how do we get there? Plus, is lack of security really just the result of a bad design?
Semiconductor Engineering sat down to discuss power, standards and the IoT with Jerry Frenkil, director of open standards at Si2; Frank Schirrmeister, group director of product marketing of the System Development Suite at Cadence; Randy Smith, vice president of marketing at Sonics; and Vojin Zivojnovic, co-founder and CEO of Aggios. What follows are excerpts of that discussion, which was held in front of a live audience at the IEEE SA Symposium on EDA and IP Interoperability in San Jose. To view part one, click here.
SE: How much more power can we save in the future compared with what we’re doing today?
Smith: More than half.
Schirrmeister: Given that we are engineers, we sometimes getting lost in optimizing the noise. One reason we go through all this trouble is that we want to optimize the time you can go between charges. I never watch a video on my game console because it’s about 100 times more power inefficient than a dedicated device. From a global power perspective, that’s really the noise. It’s still important. You need to deal with the thermal effects. You need to use the right protocol to connect the device. Which wireless interface and which decoder to you use? That’s more important holistically.
Frenkil: It’s hard to answer with a specific number, but half sounds about right. There are a couple of examples that lead me to believe we can do better. For a systems designer and a product designer, what are you concerned about? If I’m designing an IoT device, I’m not necessarily concerned about the power consumption in a data center. I may divide my architecture into different points, depending upon what I want to hit. If I want to do a really low-power node, I might do two things. One is I might do minimal processing there. Maybe I’m shipping symbols instead of data, so the data center can interpret those symbols and do all the processing. The second thing is if I’m doing minimal processing on the node, maybe I can go to a really low voltage. Today, practically speaking, not many designs are done using near-threshold voltages. It’s too hard and the performance drops off rather rapidly. If I combine judicious partitioning of my system, which allows me to do minimal local processing so I can use a much lower-voltage design, then I can go a lot lower.
Schirrmeister: I charge my phone every day. But if you look at the power envelope of this phone versus the previous generations, it seems to run for 1 day or 1.5 days. But it does so much more now than what it did five or six years ago. The question isn’t power or energy itself. It’s the workload for a given amount of power, and there we have improved much more than half. The power envelope is only going slightly down, but the amount of stuff you can do with it is much more.
SE: If you get much more efficiency out of something designed for a specific function, do IoT devices have to be very specific in what they do? And how does that change the overall design?
Schirrmeister: Absolutely that needs to happen. My first phone didn’t have the dedicated hardware decoder for video. I was always afraid when I went to the gym whether I would be able to get through my favorite show. Now you only use 10% of the power because they’ve made it a dedicated function that has been optimized for that purpose.
Zivojnovic: In mobile devices, just on the software side you can squeeze out 40% more efficiency. On devices like set-top boxes, TVs and game consoles, it’s easy to reach 80% just by tuning the software. The key is something just built for a certain purpose. It only has to do certain jobs. You can get orders of magnitude improvement because some transistors don’t have to clock. What we’re doing with clock gating and voltage gating is simply mimicking the logic. You can look at every research book or introduction to digital design and you will see there are synchronous clocks, synchronous logic and asynchronous logic. So you can go from one day to 10 days between charges. We are playing with what we call software-defined power management.
Smith: Sometimes we get caught up with what happens in the applications process with some of the sophisticated techniques we’re talking about. Our rather informal survey of customers says the average is three or less power domains on chips because it’s just too hard to design and verify if they put in a lot more than that. That’s why we need all these models and standards to build the design, analysis and verification tools. They’re not doing it now. They’re doing it on chips where they have the volume for them, but the vast majority use three or less domains. At the same time, we’re seeing designs with 120 or more blocks. There is a lot of opportunity to get more granular with controlling power, but they’re too complex using today’s IP models and tools. That’s our challenge to solve to really bring down the power.
SE: One of the hidden taxes in the IoT is security. Security costs power and cycles somewhere. Is there a better way to do it?
Zivojnovic: Security very often is an excuse for not doing a good design. We have seen cases where people don’t want to put the additional effort into a device like a set-top box. Then, when they are forced by the manufacturer, or occasionally by the regulators, most of them give up and add the mobile processor, more efficient memory or green memory. But with security you need to be constantly in charge of the unit. It’s very important. But there are better ways, such as having hardware accelerators to engage with dedicated hardware.
Smith: There is active security, and there is designed-in security, which is passive. Active includes things like encryption. It takes more power and you have to account for that. You can design in better stuff like hardware firewalls and set up your router so this block can never talk to that block. Your fingerprint can never get a message out to your radio so you’ll never be able to transmit your fingerprints someplace. That’s a simple thing you can do in design. It does mean extra effort in design, though, because you’re stopping things you don’t want to happen from ever happening. And then those in a general-purpose system, you pay the cost for encryption or security you have to enforce. There will be costs, but we can be smart about it.
Frenkil: On the encryption side, there will be some more costs but it won’t be that high. We can have a standard for encryption for the IoT. That can be built as a piece of hard IP. It will be small and power-efficient. It will consume some power, but if it’s hard IP it can be made quite power-efficient. And it’s something you can put in almost as a no-brainer. But we do need standard protocols with standard security in those protocols to enable people to define what is that encryption, what does that security processor look like.
Schirrmeister: It depends on the design. If the requirements at the end dictate that security is more important than power or the lifetime of the battery, then you have to do it. You don’t want your health information transmitted or for someone to be able to hack into it. It’s a question of what type of device you’re looking at.