FETA: a Flexible Low-Power AI/ML Accelerator for Time Series Signals

Abstract
The large variety of data that is acquired by sensors on mobile, wearable, and IoT devices has enabled numerous new applications such as long-term medical monitoring, fitness tracking, and voice control. ML algorithms such as neural networks (NNs) are often used for processing time-dependent sensor data (time-series) from these sensors. However exploitation in edge devices is still limited by the inefficient processing of the vast amounts of sensor data.Today, very few portable devices embed ML features and the rare ML tasks that are performed are often limited. Devices with tight power budgets are rarely enabled with ML functionalities at all. The numerous operations required by ML algorithms are in fact typically offloaded to the cloud, at the cost of power-hungry radio communication, long latency, and privacy risks. Thus, the design of ultra low-power NN accelerators is key to enable ML features in any battery powered device. The development of optimized, yet flexible accelerators for NNs can unlock from 2x to 10x savings in power consumption. Thanks to the design of these circuits, the execution of computing-intensive algorithms can be made possible for any portable device and create unprecedented use-cases for edge devices.
Publication Reference
FETA: a Flexible Low-Power AI/ML Accelerator for Time Series Signals, ASICS for Edge, EMERY Stéphane
Year
2024
Sponsors