Aspinity Tackles Analog Variability

Aspinity’s AML100 sensor inference chip uses analog circuits for low power, acting as an always-on block that can wake a higher-power processor if it detects a relevant event. But analog approaches have so far been fraught owing to process and environmental variations. From recent disclosures, MPR learned more about the chip and how the startup aims to overcome the kinds of variation that have plagued other analog AI efforts.

Founded in 2015 and led by CEO Tom Doyle, CTO Brandon Rumberg, and chief science officer David Graham, Aspinity raised another $5 million this autumn in a Series B round, bringing the funding total to $19 million. A new investor, automotive semiconductor specialist Unitrontech, joins earlier investor Amazon as a strategic partner for its technology.

The company’s AnalogML technology processes time-series data such as audio using analog techniques for always-on operation at less than 40 μW for common applications. Analog circuits perform feature extraction and inference, converting inference results to digital for possible further processing in an embedded FPGA and for triggering a wakeup in an external processor. The AML100 achieves 400 million operations per second (MOPS) at 40 μW power for power efficiency of 10 MOPS/W measured on customer designs.

Aspinity named its general analog platform Reconfigurable Analog Modular Platform (RAMP). Its first implementation as a neural engine is the AnalogML core that forms the basis of the AML100. The AML100 will enter full production Q1 2024 priced in the $1–$2 range for million-unit quantities; new design starts are open to any interested customer.

Aspinity’s primary efforts today focus on marketing the AML100 and expanding its use into additional markets, including automotive. Other new markets, like those requiring imaging and radar inputs, are under consideration for a new chip. The company is also considering permitting designers to license IP for inclusion in SoCs, although it won’t promote that option widely.