SkipNZ: Non-zero Value Skipping for Efficient CNN Acceleration

Joonyup Kwon; Jinhyeok Choi; Ngoc Son Pham; Sangwon Shin; Taeweon Suh

doi:10.1007/978-3-031-99857-7_4

SkipNZ: Non-zero Value Skipping for Efficient CNN Acceleration

Joonyup Kwon
, Jinhyeok Choi
, Ngoc Son Pham
, Sangwon Shin
, Taeweon Suh^*

^*Corresponding author for this work

Department of Computer Science and Engineering

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

Abstract

This paper proposes SkipNZ, a novel approach to reduce computational demands with negligible accuracy loss in the CNN inference processing. SkipNZ extends existing zero-value skipping technique and enables the skipping of unnecessary multiplications. The main idea is to filter out non-zero values if the exponent difference is large enough, so that unnecessary multiplications are skipped. The evaluation results show that the proposed technique significantly reduces the number of multiplications with negligible accuracy loss. Compared to the baseline, SkipNZ with Gap9 reduces execution time to 0.71× in AlexNet with 0.1% accuracy loss. In VGG16, SkipNZ with Gap8 lowers the execution time to 0.78× with no accuracy loss. Synthesis results confirm the practicality of the proposed approach, showing that the area and power consumption overheads of SkipNZ are only 0.5% and 0.1%, respectively, compared to the baseline.

Original language	English
Title of host publication	Euro-Par 2025
Subtitle of host publication	Parallel Processing - 31st European Conference on Parallel and Distributed Processing, Proceedings
Editors	Wolfgang E. Nagel, Diana Goehringer, Pedro C. Diniz
Publisher	Springer Science and Business Media Deutschland GmbH
Pages	48-59
Number of pages	12
ISBN (Print)	9783031998560
DOIs	https://doi.org/10.1007/978-3-031-99857-7_4
Publication status	Published - 2026
Event	31st International Conference on Parallel and Distributed Computing, Euro-Par 2025 - Dresden, Germany Duration: 2025 Aug 25 → 2025 Aug 29

Publication series

Name	Lecture Notes in Computer Science
Volume	15901 LNCS
ISSN (Print)	0302-9743
ISSN (Electronic)	1611-3349

Conference

Conference	31st International Conference on Parallel and Distributed Computing, Euro-Par 2025
Country/Territory	Germany
City	Dresden
Period	25/8/25 → 25/8/29

Bibliographical note

Publisher Copyright:
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2026.

Keywords

AI Accelerators
Convolutional Neural Networks (CNN)
Hardware Acceleration
Non-Zero value skipping

ASJC Scopus subject areas

Theoretical Computer Science
General Computer Science

Access to Document

10.1007/978-3-031-99857-7_4

Cite this

Kwon, J., Choi, J., Pham, N. S., Shin, S., & Suh, T. (2026). SkipNZ: Non-zero Value Skipping for Efficient CNN Acceleration. In W. E. Nagel, D. Goehringer, & P. C. Diniz (Eds.), Euro-Par 2025: Parallel Processing - 31st European Conference on Parallel and Distributed Processing, Proceedings (pp. 48-59). (Lecture Notes in Computer Science; Vol. 15901 LNCS). Springer Science and Business Media Deutschland GmbH. https://doi.org/10.1007/978-3-031-99857-7_4

SkipNZ: Non-zero Value Skipping for Efficient CNN Acceleration. / Kwon, Joonyup; Choi, Jinhyeok; Pham, Ngoc Son et al.
Euro-Par 2025: Parallel Processing - 31st European Conference on Parallel and Distributed Processing, Proceedings. ed. / Wolfgang E. Nagel; Diana Goehringer; Pedro C. Diniz. Springer Science and Business Media Deutschland GmbH, 2026. p. 48-59 (Lecture Notes in Computer Science; Vol. 15901 LNCS).

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

Kwon, J, Choi, J, Pham, NS, Shin, S & Suh, T 2026, SkipNZ: Non-zero Value Skipping for Efficient CNN Acceleration. in WE Nagel, D Goehringer & PC Diniz (eds), Euro-Par 2025: Parallel Processing - 31st European Conference on Parallel and Distributed Processing, Proceedings. Lecture Notes in Computer Science, vol. 15901 LNCS, Springer Science and Business Media Deutschland GmbH, pp. 48-59, 31st International Conference on Parallel and Distributed Computing, Euro-Par 2025, Dresden, Germany, 25/8/25. https://doi.org/10.1007/978-3-031-99857-7_4

Kwon J, Choi J, Pham NS, Shin S, Suh T. SkipNZ: Non-zero Value Skipping for Efficient CNN Acceleration. In Nagel WE, Goehringer D, Diniz PC, editors, Euro-Par 2025: Parallel Processing - 31st European Conference on Parallel and Distributed Processing, Proceedings. Springer Science and Business Media Deutschland GmbH. 2026. p. 48-59. (Lecture Notes in Computer Science). doi: 10.1007/978-3-031-99857-7_4

Kwon, Joonyup ; Choi, Jinhyeok ; Pham, Ngoc Son et al. / SkipNZ : Non-zero Value Skipping for Efficient CNN Acceleration. Euro-Par 2025: Parallel Processing - 31st European Conference on Parallel and Distributed Processing, Proceedings. editor / Wolfgang E. Nagel ; Diana Goehringer ; Pedro C. Diniz. Springer Science and Business Media Deutschland GmbH, 2026. pp. 48-59 (Lecture Notes in Computer Science).

@inproceedings{f06822ef6a6945f298fd6faff10ca49b,

title = "SkipNZ: Non-zero Value Skipping for Efficient CNN Acceleration",

abstract = "This paper proposes SkipNZ, a novel approach to reduce computational demands with negligible accuracy loss in the CNN inference processing. SkipNZ extends existing zero-value skipping technique and enables the skipping of unnecessary multiplications. The main idea is to filter out non-zero values if the exponent difference is large enough, so that unnecessary multiplications are skipped. The evaluation results show that the proposed technique significantly reduces the number of multiplications with negligible accuracy loss. Compared to the baseline, SkipNZ with Gap9 reduces execution time to 0.71× in AlexNet with 0.1\% accuracy loss. In VGG16, SkipNZ with Gap8 lowers the execution time to 0.78× with no accuracy loss. Synthesis results confirm the practicality of the proposed approach, showing that the area and power consumption overheads of SkipNZ are only 0.5\% and 0.1\%, respectively, compared to the baseline.",

keywords = "AI Accelerators, Convolutional Neural Networks (CNN), Hardware Acceleration, Non-Zero value skipping",

author = "Joonyup Kwon and Jinhyeok Choi and Pham, \{Ngoc Son\} and Sangwon Shin and Taeweon Suh",

note = "Publisher Copyright: {\textcopyright} The Author(s), under exclusive license to Springer Nature Switzerland AG 2026.; 31st International Conference on Parallel and Distributed Computing, Euro-Par 2025 ; Conference date: 25-08-2025 Through 29-08-2025",

year = "2026",

doi = "10.1007/978-3-031-99857-7\_4",

language = "English",

isbn = "9783031998560",

series = "Lecture Notes in Computer Science",

publisher = "Springer Science and Business Media Deutschland GmbH",

pages = "48--59",

editor = "Nagel, \{Wolfgang E.\} and Diana Goehringer and Diniz, \{Pedro C.\}",

booktitle = "Euro-Par 2025",

address = "Germany",

}

TY - GEN

T1 - SkipNZ

T2 - 31st International Conference on Parallel and Distributed Computing, Euro-Par 2025

AU - Kwon, Joonyup

AU - Choi, Jinhyeok

AU - Pham, Ngoc Son

AU - Shin, Sangwon

AU - Suh, Taeweon

PY - 2026

Y1 - 2026

N2 - This paper proposes SkipNZ, a novel approach to reduce computational demands with negligible accuracy loss in the CNN inference processing. SkipNZ extends existing zero-value skipping technique and enables the skipping of unnecessary multiplications. The main idea is to filter out non-zero values if the exponent difference is large enough, so that unnecessary multiplications are skipped. The evaluation results show that the proposed technique significantly reduces the number of multiplications with negligible accuracy loss. Compared to the baseline, SkipNZ with Gap9 reduces execution time to 0.71× in AlexNet with 0.1% accuracy loss. In VGG16, SkipNZ with Gap8 lowers the execution time to 0.78× with no accuracy loss. Synthesis results confirm the practicality of the proposed approach, showing that the area and power consumption overheads of SkipNZ are only 0.5% and 0.1%, respectively, compared to the baseline.

AB - This paper proposes SkipNZ, a novel approach to reduce computational demands with negligible accuracy loss in the CNN inference processing. SkipNZ extends existing zero-value skipping technique and enables the skipping of unnecessary multiplications. The main idea is to filter out non-zero values if the exponent difference is large enough, so that unnecessary multiplications are skipped. The evaluation results show that the proposed technique significantly reduces the number of multiplications with negligible accuracy loss. Compared to the baseline, SkipNZ with Gap9 reduces execution time to 0.71× in AlexNet with 0.1% accuracy loss. In VGG16, SkipNZ with Gap8 lowers the execution time to 0.78× with no accuracy loss. Synthesis results confirm the practicality of the proposed approach, showing that the area and power consumption overheads of SkipNZ are only 0.5% and 0.1%, respectively, compared to the baseline.

KW - AI Accelerators

KW - Convolutional Neural Networks (CNN)

KW - Hardware Acceleration

KW - Non-Zero value skipping

UR - https://www.scopus.com/pages/publications/105014491973

U2 - 10.1007/978-3-031-99857-7_4

DO - 10.1007/978-3-031-99857-7_4

M3 - Conference contribution

AN - SCOPUS:105014491973

SN - 9783031998560

T3 - Lecture Notes in Computer Science

SP - 48

EP - 59

BT - Euro-Par 2025

A2 - Nagel, Wolfgang E.

A2 - Goehringer, Diana

A2 - Diniz, Pedro C.

PB - Springer Science and Business Media Deutschland GmbH

Y2 - 25 August 2025 through 29 August 2025

ER -