Getting a Grip on the Internet of Things

QuickLogic’s CTO Tim Saxe gave a keynote Getting a Grip on the Internet of Things at the IoT Summit last week.

He started by relating how things have changed over the last 3 years when he talks to customers.

• Three years ago it was sensor hubs in smartphones and the power budget was 3mW (so one day between re-charging, something we are all well-trained to do).
• Two years ago it was sensor-hubs with a power budget of 500uW, almost an order of magnitude lower (one month between charges).
• Then IoT came along and we dropped almost another order of magnitude with enterprise wearables wanting 80uW power budgets, which will last of 6 months or is low enough to make the various energy harvesting approaches workable (so battery life becomes effectively infinite).

In the past systems were primarily built around software. If the processor wasn’t fast enough then up the frequency, if that can’t be done then add more cores, if that can’t be done then add a big FPGA to accelerate some algorithms. This approach is very power hungry and IoT turns everything upside down with the extremely limited power budgets.

Let’s look at what 80uW allows you:

• accelerometer takes 14uW
• BlueTooth Smart takes 12uW
• power management takes 20uW
• which leaves 34uW for processing

A representative microprocessor takes about 100uA/MHz so you can afford around ⅓MIPS.

BTW a pet peeve of mine. If your processor runs 333K instructions per second then it is ⅓MIPS, not ⅓MIP. The S in MIPS is not making it plural, it is the “second” of “per second”. End of pet peeve.

A basic pedometer takes about ⅓MIPS but anything with more sophistication needs more. Dedicated hardware is too complex to build high-level decision-making on top of. But pure software is too energy intensive. What is required is to move the few energy intensive parts to hardware and everywhere else keep the flexibility of software to get both accuracy and low-power.

Accuracy turns out to be a nebulous concept because more accurate measurements take more power. In fact like the three most important factors in real estate all being location, Tim repeated this three times. But in practice more power means more data lost when the battery needs replenishing, meaning that the power-hungry very accurate pedometer may be outclassed by a less accurate but less power-hungry approach.

A solution to a lot of complicated issues like natural language processing is to use machine learning. This isn’t always appropriate: machine learning is going to kill a lot of pilots and destroy a lot of money before it learns how to fly a plane. But for non-critical applications it is often a superb way to build an application that soon outclasses even the best hand-constructed models.

So Tim wrapped up by going back to reiterate that the IoT requires a change in how we should think about designs. The two big takeaways are:
[LIST=1]