INT Forschung: Energieeffizienter analoger Multiplizierer und Spannungs-/Zeitumsetzer für neuronale Netze

27. Juni 2022 / INT

Das INT stellt ein Paper und ein Poster auf der 20th IEEE International New Circuits and Systems Conference (NEWCAS 2022) vor.
[Bild: www.newcas2022.org]

Design of an Energy Efficient Analog Two-Quadrant Multiplier Cell Operating in Weak Inversion

Raphael Nägele, Jakob Finkbeiner, Markus Grözing, Manfred Berroth

Abstract — Analog low precision arithmetic circuits offer a significantly higher energy efficiency than their digital counterparts which makes them ideally suited for low precision neuromorphic processing circuits. An analog two-quadrant multiplier cell consisting of only two MOSFETs with multi-bit resolution is presented. It operates in weak inversion with the backgate used as multiplicator input consuming less than 1 fJ per operation. A 22 nm FD-SOI CMOS technology is used for simulations. 

Projektseite

Design of an Energy Efficient Voltage-to-Time Converter with Rectified Linear Unit Characteristics for Artificial Neural Networks

Jakob Finkbeiner, Raphael Nägele, Markus Grözing, Manfred Berroth

Abstract — Machine learning at the edge is fast, secure and robust. Because of the limited power budget, calculations need to be very energy efficient. This paper presents the design of an energy efficient voltage-to-time converter circuit in 22nm FD-SOI CMOS technology. It has a rectified linear unit transfer characteristic and is suited for analog mixed signal computing architectures for artificial neural network inference. Depending on whether a calibration process is performed or not, the resolution for a maximum pulse width of 430 ps is 3.0 b or 6.4 b. The energy consumption per cycle stays below 3 fJ.

Projektseite

Zum Seitenanfang