TY - GEN
T1 - Multiple approximate instances in neural processing units for energy-efficient circuit synthesis
T2 - 2021 International Conference on Compilers, Architectures, and Synthesis for Embedded Systems, CASES 2021
AU - Alan, Tanfer
AU - Castro-Godínez, Jorge
AU - Henkel, Jörg
N1 - Publisher Copyright:
© 2021 ACM.
PY - 2021/9/30
Y1 - 2021/9/30
N2 - We present an architectural approach toward energy-efficient synthesis of circuits used in neural processing units. Neural network applications are shown to tolerate varying operand precisions between different inputs, accuracy targets, their phases, and learning methods, without significantly impacting the classification accuracy. Using multiple instances of systolic arrays at different precisions, we show that significant energy gains are possible beyond the conventional approach, using the same circuit for all precisions.
AB - We present an architectural approach toward energy-efficient synthesis of circuits used in neural processing units. Neural network applications are shown to tolerate varying operand precisions between different inputs, accuracy targets, their phases, and learning methods, without significantly impacting the classification accuracy. Using multiple instances of systolic arrays at different precisions, we show that significant energy gains are possible beyond the conventional approach, using the same circuit for all precisions.
UR - http://www.scopus.com/inward/record.url?scp=85117494608&partnerID=8YFLogxK
U2 - 10.1145/3451939.3477594
DO - 10.1145/3451939.3477594
M3 - Contribución a la conferencia
AN - SCOPUS:85117494608
T3 - Proceedings - 2021 International Conference on Compilers, Architectures, and Synthesis for Embedded Systems, CASES 2021
SP - 3
EP - 5
BT - Proceedings - 2021 International Conference on Compilers, Architectures, and Synthesis for Embedded Systems, CASES 2021
PB - Association for Computing Machinery, Inc
Y2 - 8 October 2021 through 15 October 2021
ER -