A dot product kernel using rapidly switched analog circuit

In a world driven by technology and hand-held devices, there is ubiquitous demand for high-performance, low-energy processing engines. In this thesis, we present rapidly switched analog circuit (RSAC), a new circuit architecture, to implement an energy-efficient mixed-signal dot product (DP) kernel for machine learning and signal processing applications. RSAC operates by fast switching the analog inputs to the output via variable width digital pulses. A description of the different components of RSAC, along with a detailed accuracy and energy consumption analysis is presented. We show two RSAC designs that span the different design options and technology nodes. Simulations for the first design in a 130 nm process show energy savings of 19x to 32x compared to a digital implementation for signal-to-quantization-noise ratios (SQNRs) of 30 dB to 24 dB, respectively. Simulations for the second design in a 28 nm FDSOI process show energy savings of 15.7x, 4x, 2.1x compared to a digital implementation running at the same sampling frequency for SQNRs of 8 dB, 14 dB and 20 dB, respectively. Finally, we present the design of an emotion recognition system composed solely of SAC-based dot-products. Based on the behavioral and energy models developed in this thesis, we obtain energy savings of 45% and 49% compared to a digital implementation for average probabilities of error of 0.23 and 0.07, running at frequencies of 1.87 MHz and 1.7 MHz, respectively.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Electrical & Computer Engineering

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2015

Author dc:creator

• Nahlus, Ihab

Contributors dc:contributor

• Shanbhag, Naresh R.

• low energy
• mixed-signal
• dot product

Statement dc:rights

• Copyright 2015 Ihab Nahlus

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/88010

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/88010