Speaking at the European MEMS and Sensors summit organized by SEMI, Wiegand emphasized that intelligence not only has to be at the edge – for reasons of power consumption, latency and privacy – but also at the edge of the equipment.
Wiegand gave the example of the BMA400 MEMS accelerometer with built-in step counter that can be the only device that is on in a piece of equipment, consuming just 1 microamp and yet is able to recognize a significant movement and wake up the rest of the device.
Wiegand’s second example was the BHI260, a six-axis IMU that includes a 32bit Fuser2 processor based on the ARC microprocessor available from Synopys (see Bosch updates smart sensor hubs). This core can run at 20MHz clock frequency while drawing 950 microamps and has been used to host a machine learning to help equipment adapt to individual users, Wiegand said.
Speaking to eeNews Europe after his presentation Wiegand said that Bosch is interested in deeply embedded AI within its sensors to minimize data transfers and energy consumption. “For embedded AI you cannot use a Google TPU [tensor processing unit] for wearables,” he said. “Using the ASIC in our current device [the BHI260] we have made a proof of concept. We have taken ML and run it on such a smart microcontroller.”
When asked whether machine learning hardware would replace a software approach and whether that would be licensed in, like the ARC intellectual property, or developed in house Wiegand said: That decision has not been made yet. It won’t be in the next generation of MEMS sensor. There is still a lot of work we can do with algorithms.”
While dedicated machine learning hardware produces the lowest power consumption Bosch is still valuing the flexibility of software running on a general-purpose microcontroller core within its ASIC.
Wiegand also said in his talk that energy harvesting is not something that Bosch Sensortec would be pursuing in the near future. “We have not seen the cool solution yet. For now the solution is still a battery.”
Related links and articles: