FlashMAC: An Energy-Efficient Analog-Digital Hybrid MAC with Variable Latency-Aware Scheduling
Abstract
The demand for energy-efficient DNN accelerator on edge devices is rapidly increasing. Diverse approaches using analog circuits to overcome the limitations of digital logics – leakage and limited power scaling (∼CV2DD) - have been developed. Previous works on analog MACs can be categorized according to their operation domains. The voltage domain MAC have been studied first in the field of matrix multiplier for CNN application [1]. However, it suffered from PVT variation and limited V DD scaling of analog circuits. The frequency-domain MAC [2] took advantage of low supply voltage and high throughput, but a large power consumption of the internal oscillator and its nonlinearity restrict its accuracy of computing. A delay-based MAC approach in the time-domain [3] –[5] recently drew attention because of its easy integration with digital circuits, low power consumption, and small area. However, their support for higher precision is limited and needs multiple delay lines with digital adders for precision scaling. Show more
Publication status
publishedExternal links
Book title
2021 IEEE Asian Solid-State Circuits Conference (A-SSCC)Pages / Article No.
Publisher
IEEEEvent
Organisational unit
02533 - Institut für Neuroinformatik / Institute of Neuroinformatics
More
Show all metadata
ETH Bibliography
yes
Altmetrics