I’ve been following Eta Compute for a while. Like some other vendors, their interest is putting a little machine learning capability at the “edge” – in a microcontroller right at the sensors. What makes them a bit different is their focus on doing this at extremely low power levels.Eta’s new ECM3532 is an MCU with a Cortex M3 and DSP with dual MACs. The latter, of course, is optimized for multiply-accumulate operations, which are at the heart of inferencing. The company claims (I’ve seen their eval boards running off 1 cm2 solar cells) power consumption of under 5 µA/MHz, which is around an order of magnitude better than most ultra-low-power MCUs. Instead of dynamic voltage and frequency scaling, they tout continuous voltage and frequency scaling, with a near-threshold low-end Vdd of 0.54 V.Near- and sub-threshold operation yields very low power operation at low frequencies. Ambiq is one of the pioneers in this area. FETs can operate in three regions: subthreshold, linear, and saturated. But in the subthreshold region leakage is a problem, and temperature even more so. Transistors are not well characterized at those voltage levels, so vendors resort to clever, often secret, tricks. I’m told digital watches, at least the non-smart versions, operate in the subthreshold region giving them tremendous battery lives. For more on subthreshold voltage operation of FETs see this.One of those tricks Eta uses is a non-synchronous architecture. Though there is a clock, the logic isn’t conventionally clocked. It’s self-timed. Synchronous circuits need plenty of margin to insure all of the timing is properly closed; self-timed versions report back as soon as an operation has completed so the next can begin. Absent this, since FETs operating in or near the subthreshold region can have wildly-varying characteristics depending on temperature, huge setup and hold margins would have to be designed in.Why would anyone want an inferencing engine at the edge? Potential uses include listening for wake-up words (think “Hey Google”), sensing glass breaking, and even a bit of image recognition. Local processing means less bandwidth to the cloud, and a lot less power consumption as comm protocols can drain batteries quickly.Samples of the ECM3532 are available today. In production volumes they will be under $5.Note: This section is about something I personally find cool, interesting or important and want to pass along to readers. It is not influenced by vendors.