Ambiq was founded in 2010 as a spin off from University of Michigan to commercialize sub-threshold voltage operation of circuits, which it has applied to families of real-time clock circuits and its Apollo1 and Apollo2 families of ARM-based microcontrollers. The Apollo microcontrollers can operate at voltages below 0.5V and the company claims that this can provide a 10-fold improvement in MCU power consumption compared with competitors’ MCUs. This is mainly because power consumption scales with the square of voltage.
The company has made a particular point of addressing wearables markets where low power consumption for battery operation is a priority and, as a new market, their are fewer issues over established players using legacy to maintain position (see CEO interview: Ambiq sees broader options for low voltage). This has also exposed Ambiq to a number of customers working on machine learning applications in things like speech interfaces and keyword detection, Scott Hanson, founder and CTO of Ambiq, told eeNews Europe in a telephone interview.
“Apollo is already being used for neural networks through CMSIS libraries for Cortex-M4. It turns out sub-threshold [voltage operation] plus ARM is a good platform for machine learning,” Hanson said. “We have voice solutions running on Apollo. We have partnered with DSP Concepts Inc. on beamforming and noise reduction and with Sensory Inc. on keyword detection.”
In a YouTube-published webinar Hanson said that Ambiq’s customers today are being limited to the use of relatively lightweight neural networks. “Neural networks can very quickly result in model sizes that are enormous and that can be a pretty big compute burden. The Cortex-M4 devices that we have today and that are prevalent in wearables are really not tuned for neural networks,” he said (see Valencell-Ambiq webinar). Hanson continued: “There is a need for something more. It’s an opportunity and also the next great energy problem to solve.”
Next: On the phone
In the interview with eeNews Europe Hanson added: “We are developing a next-generation product working with existing and new customer influences. It is intended to address all workloads including neural networks and non neural networks.”
He made the point that to do something like keyword detection an SoC needs to be able to handle a variety of workloads including sensor data fusion, pre-processing, noise reduction and power management. It takes a well-crafted combination of hardware and software and you may also need to be capable of inferencing in a variety of neural network styles.
Hanson declined to say whether Ambiq’s upcoming processing platform would be based on a licensed-in architecture from ARM – as Apollo is, based on an architecture licensed in from someone else, or whether it would be a home-grown solution.
Building out from the increasingly popular open-standard RISC-V ISA could provide an approach to a RISC-plus-NN-accelerator architecture but Hanson declined to comment on that possibility. Hanson said Ambiq is not ready to announce the solution but it would form the basis of a new family of products built for neural networks and other types of workload.
“In deeply embedded systems you have to be able to do sensor data pre-processing, inference, communications, power management” said Hanson. He added that there is also then the wide variety of applications and machine learning approaches to those applications, of optimal data and weight resolutions, and of the design frameworks and tool environments.,
“I am a big believer in configurability and big believer in being able to plug into an ecosystem,” Hanson said. “We do need to be able to support a wide range of networks. We see interest in binarized networks, in 4- and 8bit integer, in higher-resolution variable-precision data, and in variable precision in weights across a network. It’s important to take all that and provide a flexible architecture.
Next: How soon?
When asked how soon Ambiq would be able to offer its next generation processing platform Hanson said: “We’re deep in the design phase and we have customer engagements. We’ll be getting first spin of the product soon.”
This suggests to eeNews Europe that a formal announcement is months rather than years away and that those early customers that Hanson mentioned could benefit in 2018, even if more general availability happens in 2019.
Related links and articles: