Hartung-Gorre Verlag

Inh.: Dr. Renate Gorre

D-78465 Konstanz

Fon: +49 (0)7533 97227

Fax: +49 (0)7533 97228



Series in Microelectronics

edited by       Qiuting Huang

Andreas Schenk

Mathieu Maurice Luisier

Bernd Witzigmann

Vol. 238






Lukas Arno Jakob Cavigelli


Towards Energy-Efficient

Convolutional Neural

Network Inference


2019. XVIII, 234 pages. € 64,00.

ISBN 978-3-86628-651-1


















Deep  learning  and  particularly  convolutional  neural  networks (CNNs) have  become the  method of  choice for most  computer vision  tasks.   The  achieved  leap  in  accuracy  has  dramatically increased  the  range  of  possibilities  and  created  a  demand  for running these compute and memory intensive algorithms on embedded and mobile devices.  In this thesis, we evaluate the capabilities of software-programmable hardware, dive into specialized  accelerators,  and  explore  the  potential  of  extremely  quantized  CNNs—all  with  special  consideration  to  external  memory bandwidth, which dominates the overall energy cost.  We establish  that — including  I/O — software-programmable  platforms  can achieve  10–40 GOp/s/W,  our  specialized  accelerator  for  fixedpoint CNNs achieves 630 GOp/s/W, binary-weight CNNs can be implemented  with  up  to  5.9 TOp/s/W  and  very  small  binarized neural  networks  implementable  with  purely  combinational  logic could be run directly on the sensor with 670 TOp/s/W.



About the Author


Lukas Cavigelli was born in Zurich, Switzerland in 1990.  He received the M.Sc.  degree in electrical engineering and information  technology  from  ETH  Zurich  in  2014.   He  then  joined  the group of Prof.  Dr.  Luca Benini at the Integrated Systems Laboratory,  ETH  Zurich,  to  pursue  a  doctorate.   Since  his  graduation in 2019, he has continued his work there as a post-doctoral researcher.  He has received the best paper award at the VLSI-SoC’13 and the ICDSC’17 conferences and the Donald O. Pederson best paper award of the IEEE Trans. on CAD in 2019.


Series in Microelectronics

Direkt bestellen bei / to order directly from:

Hartung-Gorre Verlag / D-78465 Konstanz / Germany

Telefon: +49 (0) 7533 97227  Telefax: +49 (0) 7533 97228
http://www.hartung-gorre.de   eMail: verlag@hartung-gorre.de