A Light-Speed Large Language Model Accelerator with Optical Stochastic Computing

Salma Afifi; Oluwaseun Alo; Ishan Thakkar; Sudeep Pasricha

doi:10.1145/3716368.3735299

A Light-Speed Large Language Model Accelerator with Optical Stochastic Computing

Salma Afifi
, Oluwaseun Alo
, Ishan Thakkar
, Sudeep Pasricha

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

1 Scopus citations

Abstract

To address the increasingly intensive computational demands of attention-based large language models (LLMs), there is a growing interest in developing energy-efficient and high-speed hardware accelerators. To that end, photonics is being considered as an alternative technology to digital electronics. This work introduces a novel optical hardware accelerator that leverages stochastic computing principles for LLMs. Our proposed accelerator incorporates full-range optical stochastic multipliers and stochastic-analog compute-capable optical-to-electrical transducer units to efficiently handle static and dynamic tensor computations in attention-based models. Our analysis shows that our accelerator exhibits at least 7.6× speedup and 1.3× lower energy compared to state-of-the-art LLMs hardware accelerators.

Original language	English
Title of host publication	GLSVLSI 2025 - Proceedings of the Great Lakes Symposium on VLSI 2025
Pages	922-928
Number of pages	7
ISBN (Electronic)	9798400714962
DOIs	https://doi.org/10.1145/3716368.3735299
State	Published - Jun 29 2025
Event	35th Edition of the Great Lakes Symposium on VLSI 2025, GLSVLSI 2025 - New Orleans, United States Duration: Jun 30 2025 → Jul 2 2025

Publication series

Name	Proceedings of the ACM Great Lakes Symposium on VLSI, GLSVLSI

Conference

Conference	35th Edition of the Great Lakes Symposium on VLSI 2025, GLSVLSI 2025
Country/Territory	United States
City	New Orleans
Period	6/30/25 → 7/2/25

Bibliographical note

Publisher Copyright:
© 2025 Copyright held by the owner/author(s).

Keywords

inference acceleration
optical computing
silicon photonics
stochastic computing
Transformer neural networks

ASJC Scopus subject areas

General Engineering

Access to Document

10.1145/3716368.3735299

Cite this

@inproceedings{6f3c650b597c44e1a8731c32abf412a3,

title = "A Light-Speed Large Language Model Accelerator with Optical Stochastic Computing",

abstract = "To address the increasingly intensive computational demands of attention-based large language models (LLMs), there is a growing interest in developing energy-efficient and high-speed hardware accelerators. To that end, photonics is being considered as an alternative technology to digital electronics. This work introduces a novel optical hardware accelerator that leverages stochastic computing principles for LLMs. Our proposed accelerator incorporates full-range optical stochastic multipliers and stochastic-analog compute-capable optical-to-electrical transducer units to efficiently handle static and dynamic tensor computations in attention-based models. Our analysis shows that our accelerator exhibits at least 7.6× speedup and 1.3× lower energy compared to state-of-the-art LLMs hardware accelerators.",

keywords = "inference acceleration, optical computing, silicon photonics, stochastic computing, Transformer neural networks",

author = "Salma Afifi and Oluwaseun Alo and Ishan Thakkar and Sudeep Pasricha",

note = "Publisher Copyright: {\textcopyright} 2025 Copyright held by the owner/author(s).; 35th Edition of the Great Lakes Symposium on VLSI 2025, GLSVLSI 2025 ; Conference date: 30-06-2025 Through 02-07-2025",

year = "2025",

month = jun,

day = "29",

doi = "10.1145/3716368.3735299",

language = "English",

series = "Proceedings of the ACM Great Lakes Symposium on VLSI, GLSVLSI",

pages = "922--928",

booktitle = "GLSVLSI 2025 - Proceedings of the Great Lakes Symposium on VLSI 2025",

}

Afifi, S, Alo, O, Thakkar, I & Pasricha, S 2025, A Light-Speed Large Language Model Accelerator with Optical Stochastic Computing. in GLSVLSI 2025 - Proceedings of the Great Lakes Symposium on VLSI 2025. Proceedings of the ACM Great Lakes Symposium on VLSI, GLSVLSI, pp. 922-928, 35th Edition of the Great Lakes Symposium on VLSI 2025, GLSVLSI 2025, New Orleans, United States, 6/30/25. https://doi.org/10.1145/3716368.3735299

A Light-Speed Large Language Model Accelerator with Optical Stochastic Computing. / Afifi, Salma; Alo, Oluwaseun; Thakkar, Ishan et al.
GLSVLSI 2025 - Proceedings of the Great Lakes Symposium on VLSI 2025. 2025. p. 922-928 (Proceedings of the ACM Great Lakes Symposium on VLSI, GLSVLSI).

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

TY - GEN

T1 - A Light-Speed Large Language Model Accelerator with Optical Stochastic Computing

AU - Afifi, Salma

AU - Alo, Oluwaseun

AU - Thakkar, Ishan

AU - Pasricha, Sudeep

PY - 2025/6/29

Y1 - 2025/6/29

N2 - To address the increasingly intensive computational demands of attention-based large language models (LLMs), there is a growing interest in developing energy-efficient and high-speed hardware accelerators. To that end, photonics is being considered as an alternative technology to digital electronics. This work introduces a novel optical hardware accelerator that leverages stochastic computing principles for LLMs. Our proposed accelerator incorporates full-range optical stochastic multipliers and stochastic-analog compute-capable optical-to-electrical transducer units to efficiently handle static and dynamic tensor computations in attention-based models. Our analysis shows that our accelerator exhibits at least 7.6× speedup and 1.3× lower energy compared to state-of-the-art LLMs hardware accelerators.

AB - To address the increasingly intensive computational demands of attention-based large language models (LLMs), there is a growing interest in developing energy-efficient and high-speed hardware accelerators. To that end, photonics is being considered as an alternative technology to digital electronics. This work introduces a novel optical hardware accelerator that leverages stochastic computing principles for LLMs. Our proposed accelerator incorporates full-range optical stochastic multipliers and stochastic-analog compute-capable optical-to-electrical transducer units to efficiently handle static and dynamic tensor computations in attention-based models. Our analysis shows that our accelerator exhibits at least 7.6× speedup and 1.3× lower energy compared to state-of-the-art LLMs hardware accelerators.

KW - inference acceleration

KW - optical computing

KW - silicon photonics

KW - stochastic computing

KW - Transformer neural networks

UR - https://www.scopus.com/pages/publications/105017807467

UR - https://www.scopus.com/pages/publications/105017807467#tab=citedBy

U2 - 10.1145/3716368.3735299

DO - 10.1145/3716368.3735299

M3 - Conference contribution

AN - SCOPUS:105017807467

T3 - Proceedings of the ACM Great Lakes Symposium on VLSI, GLSVLSI

SP - 922

EP - 928

BT - GLSVLSI 2025 - Proceedings of the Great Lakes Symposium on VLSI 2025

T2 - 35th Edition of the Great Lakes Symposium on VLSI 2025, GLSVLSI 2025

Y2 - 30 June 2025 through 2 July 2025

ER -

A Light-Speed Large Language Model Accelerator with Optical Stochastic Computing

Abstract

Publication series

Conference

Bibliographical note

Keywords

ASJC Scopus subject areas

Access to Document

Other files and links

Fingerprint

Cite this