RePIM: Joint Exploitation of Activation and Weight Repetitions for In-ReRAM DNN Acceleration

Chen Yang Tsai; Chin Fu Nien; Tz Ching Yu; Hung Yu Yeh; Hsiang Yun Cheng

doi:10.1109/DAC18074.2021.9586315

RePIM: Joint Exploitation of Activation and Weight Repetitions for In-ReRAM DNN Acceleration

Chen Yang Tsai, Chin Fu Nien, Tz Ching Yu, Hung Yu Yeh, Hsiang Yun Cheng

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review

14 Scopus citations

Abstract

Eliminating redundant computations is a common approach to improve the performance of ReRAM-based DNN accelerators. While existing practical ReRAM-based accelerators eliminate part of the redundant computations by exploiting sparsity in inputs and weights or utilizing weight patterns of DNN models, they fail to identify all the redundancy, resulting in many unnecessary computations. Thus, we propose a practical design, RePIM, that is the first to jointly exploit the repetition of both inputs and weights. Our evaluation shows that RePIM is effective in eliminating unnecessary computations, achieving an average of 15.24× speedup and 96.07% energy savings over the state-of-the-art practical ReRAM-based accelerator.

Original language	English
Title of host publication	2021 58th ACM/IEEE Design Automation Conference, DAC 2021
Publisher	Institute of Electrical and Electronics Engineers Inc.
Pages	589-594
Number of pages	6
ISBN (Electronic)	9781665432740
DOIs	https://doi.org/10.1109/DAC18074.2021.9586315
State	Published - 05 12 2021
Externally published	Yes
Event	58th ACM/IEEE Design Automation Conference, DAC 2021 - San Francisco, United States Duration: 05 12 2021 → 09 12 2021

Publication series

Name	Proceedings - Design Automation Conference
Volume	2021-December
ISSN (Print)	0738-100X

Conference

Conference	58th ACM/IEEE Design Automation Conference, DAC 2021
Country/Territory	United States
City	San Francisco
Period	05/12/21 → 09/12/21

Bibliographical note

Publisher Copyright:
© 2021 IEEE.

Access to Document

10.1109/DAC18074.2021.9586315

Cite this

Tsai, C. Y., Nien, C. F., Yu, T. C., Yeh, H. Y., & Cheng, H. Y. (2021). RePIM: Joint Exploitation of Activation and Weight Repetitions for In-ReRAM DNN Acceleration. In 2021 58th ACM/IEEE Design Automation Conference, DAC 2021 (pp. 589-594). (Proceedings - Design Automation Conference; Vol. 2021-December). Institute of Electrical and Electronics Engineers Inc.. https://doi.org/10.1109/DAC18074.2021.9586315

@inproceedings{92a3b564c3674d7f9376d6236cbbd41e,

title = "RePIM: Joint Exploitation of Activation and Weight Repetitions for In-ReRAM DNN Acceleration",

abstract = "Eliminating redundant computations is a common approach to improve the performance of ReRAM-based DNN accelerators. While existing practical ReRAM-based accelerators eliminate part of the redundant computations by exploiting sparsity in inputs and weights or utilizing weight patterns of DNN models, they fail to identify all the redundancy, resulting in many unnecessary computations. Thus, we propose a practical design, RePIM, that is the first to jointly exploit the repetition of both inputs and weights. Our evaluation shows that RePIM is effective in eliminating unnecessary computations, achieving an average of 15.24× speedup and 96.07% energy savings over the state-of-the-art practical ReRAM-based accelerator.",

author = "Tsai, {Chen Yang} and Nien, {Chin Fu} and Yu, {Tz Ching} and Yeh, {Hung Yu} and Cheng, {Hsiang Yun}",

note = "Publisher Copyright: {\textcopyright} 2021 IEEE.; 58th ACM/IEEE Design Automation Conference, DAC 2021 ; Conference date: 05-12-2021 Through 09-12-2021",

year = "2021",

month = dec,

day = "5",

doi = "10.1109/DAC18074.2021.9586315",

language = "英语",

series = "Proceedings - Design Automation Conference",

publisher = "Institute of Electrical and Electronics Engineers Inc.",

pages = "589--594",

booktitle = "2021 58th ACM/IEEE Design Automation Conference, DAC 2021",

address = "美国",

}

Tsai, CY, Nien, CF, Yu, TC, Yeh, HY & Cheng, HY 2021, RePIM: Joint Exploitation of Activation and Weight Repetitions for In-ReRAM DNN Acceleration. in 2021 58th ACM/IEEE Design Automation Conference, DAC 2021. Proceedings - Design Automation Conference, vol. 2021-December, Institute of Electrical and Electronics Engineers Inc., pp. 589-594, 58th ACM/IEEE Design Automation Conference, DAC 2021, San Francisco, United States, 05/12/21. https://doi.org/10.1109/DAC18074.2021.9586315

RePIM: Joint Exploitation of Activation and Weight Repetitions for In-ReRAM DNN Acceleration. / Tsai, Chen Yang; Nien, Chin Fu; Yu, Tz Ching et al.
2021 58th ACM/IEEE Design Automation Conference, DAC 2021. Institute of Electrical and Electronics Engineers Inc., 2021. p. 589-594 (Proceedings - Design Automation Conference; Vol. 2021-December).

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review

TY - GEN

T1 - RePIM

T2 - 58th ACM/IEEE Design Automation Conference, DAC 2021

AU - Tsai, Chen Yang

AU - Nien, Chin Fu

AU - Yu, Tz Ching

AU - Yeh, Hung Yu

AU - Cheng, Hsiang Yun

PY - 2021/12/5

Y1 - 2021/12/5

N2 - Eliminating redundant computations is a common approach to improve the performance of ReRAM-based DNN accelerators. While existing practical ReRAM-based accelerators eliminate part of the redundant computations by exploiting sparsity in inputs and weights or utilizing weight patterns of DNN models, they fail to identify all the redundancy, resulting in many unnecessary computations. Thus, we propose a practical design, RePIM, that is the first to jointly exploit the repetition of both inputs and weights. Our evaluation shows that RePIM is effective in eliminating unnecessary computations, achieving an average of 15.24× speedup and 96.07% energy savings over the state-of-the-art practical ReRAM-based accelerator.

AB - Eliminating redundant computations is a common approach to improve the performance of ReRAM-based DNN accelerators. While existing practical ReRAM-based accelerators eliminate part of the redundant computations by exploiting sparsity in inputs and weights or utilizing weight patterns of DNN models, they fail to identify all the redundancy, resulting in many unnecessary computations. Thus, we propose a practical design, RePIM, that is the first to jointly exploit the repetition of both inputs and weights. Our evaluation shows that RePIM is effective in eliminating unnecessary computations, achieving an average of 15.24× speedup and 96.07% energy savings over the state-of-the-art practical ReRAM-based accelerator.

UR - http://www.scopus.com/inward/record.url?scp=85119407167&partnerID=8YFLogxK

U2 - 10.1109/DAC18074.2021.9586315

DO - 10.1109/DAC18074.2021.9586315

M3 - 会议稿件

AN - SCOPUS:85119407167

T3 - Proceedings - Design Automation Conference

SP - 589

EP - 594

BT - 2021 58th ACM/IEEE Design Automation Conference, DAC 2021

PB - Institute of Electrical and Electronics Engineers Inc.

Y2 - 5 December 2021 through 9 December 2021

ER -

RePIM: Joint Exploitation of Activation and Weight Repetitions for In-ReRAM DNN Acceleration

Abstract

Publication series

Conference

Bibliographical note

Access to Document

Other files and links

Fingerprint

Cite this