Aspinity’s AML100 uses analog machine learning to extend battery life by up to 20 times

Aspinity has introduced the first member of its analogML family, AML100, the industry’s first and only micro-machine learning (ML) solution that operates entirely in the analog domain. AML100 can reduce always-on system power consumption by 95%, enabling manufacturers to significantly extend the battery life of today’s devices or migrate AC-powered always-on devices to battery – opening up entirely new product categories for voice recognition devices, smart home commercial security, predictive and preventive maintenance, and biomedical monitoring.

Minimizing the amount and movement of data in a system is one of the most effective ways to reduce power consumption, but today’s always-on devices do not have this capability. Instead, they need to monitor the environment and constantly collect and digitize large amounts of local analog data, wasting a lot of system power to process data that is barely relevant to the application. In contrast, AML100 achieves significant system-level energy savings by offloading ML workloads to ultra-low-power analog computations, where it can determine data relevance with high accuracy at near-zero power consumption. This makes AML100 the only device that can perform AI computations while the data is still analog, reducing data on the sensor and keeping digital components in low-power modes until important data is detected. This eliminates the power penalty for digitizing, digitally processing, and transmitting useless data.

“We’ve long realized that reducing the power of each chip in an always-on system would only incrementally extend battery life. For manufacturers that need revolutionary performance, that’s not good enough,” said Tom Doyle, founder and CEO of Aspinity. “The AML100 reduces always-on system power to less than 100µA, unlocking the potential for thousands of new battery-operated applications.”

AML100 interior

At the heart of the AML100 is a set of independent, configurable analog blocks (CABs) that are fully programmable in software to support a wide range of functions, including sensor interfaces and machine learning. This versatility provides a huge advantage over other analog approaches, which are rigid and only address a single function, however, the AML100 is flexible and can be reprogrammed in the field via software updates or new algorithms for other always-on applications.

The precise programmability of the AML100 analog circuitry also eliminates the chip-to-chip performance inconsistencies found in standard analog CMOS processes, which severely limit the use of highly complex analog chips.

Key features of AML100

• Power consumption is less than 20µA

• Intelligently reduces data size by up to 100x while it is still in simulation

• Field programmable to meet the needs of a variety of always-on applications

• Pre-convolutional collection using patented analog compression technology to maintain accuracy of wake word engines in voice devices

• Supports 4 analog sensors (microphone, accelerometer, etc.) in any combination

• Available in 7mm x 7mm 48-pin QFN package

Availability

Aspinity’s AML100 is currently sampling to lead customers, with production scheduled for Q4 2022. Customers can evaluate the capabilities of the AML100 by purchasing one of Aspinity’s integrated hardware-software evaluation kits: EVK1 for glass break and T3/T4 siren tone detection or EVK2 pre-convolutional collection and transmission for voice detection.