Skip to main content

Lund University Publications

LUND UNIVERSITY LIBRARIES

CPAU : a hardware efficient configurable posit arithmetic unit

Wu, Yue ; Mammadzada, Fuad ; Westring, Kristoffer LU and Rodrigues, Joachim LU (2026) 24th IEEE Interregional NEWCAS Conference, NEWCAS 2026 In 24th IEEE Interregional NEWCAS Conference, NEWCAS 2026 - Proceedings
Abstract

The Posit number system features tapered precision, providing higher accuracy for values near unity while maintaining a wide dynamic range, making the format an attractive alternative to IEEE 754 for machine-learning workloads. This work presents CPAU, a unified and configurable Posit arithmetic architecture that integrates a quire-based MAC unit and a general-purpose execution unit supporting addition, multiplication, division, exponential, and square root. CPAU employs a shared decode-execute-encode pipeline across all operators, substantially reducing hardware cost while preserving accuracy in serial computations. A dedicated exponential unit enables full softmax evaluation directly in the Posit domain without format conversions.... (More)

The Posit number system features tapered precision, providing higher accuracy for values near unity while maintaining a wide dynamic range, making the format an attractive alternative to IEEE 754 for machine-learning workloads. This work presents CPAU, a unified and configurable Posit arithmetic architecture that integrates a quire-based MAC unit and a general-purpose execution unit supporting addition, multiplication, division, exponential, and square root. CPAU employs a shared decode-execute-encode pipeline across all operators, substantially reducing hardware cost while preserving accuracy in serial computations. A dedicated exponential unit enables full softmax evaluation directly in the Posit domain without format conversions. Synthesis in a 22 nm technology demonstrates significant area improvements of more than 34% and 50% for multiplier and adder area, respectively. Moreover, the implementation of exponential function provides the functionality required for inference and training.

(Less)
Please use this url to cite or link to this publication:
author
; ; and
organization
publishing date
type
Chapter in Book/Report/Conference proceeding
publication status
published
subject
keywords
Decode, Encode, Exponential, ML training, Posit arithmetic, Softmax
host publication
24th IEEE Interregional NEWCAS Conference, NEWCAS 2026 : an IEEE CASS flagship conference. June 21-24, 2026, Saguenay, Québec, Canada: proceedings - an IEEE CASS flagship conference. June 21-24, 2026, Saguenay, Québec, Canada: proceedings
series title
24th IEEE Interregional NEWCAS Conference, NEWCAS 2026 - Proceedings
pages
5 pages
publisher
IEEE - Institute of Electrical and Electronics Engineers Inc.
conference name
24th IEEE Interregional NEWCAS Conference, NEWCAS 2026
conference location
Chicoutimi, Canada
conference dates
2026-06-21 - 2026-06-24
external identifiers
  • scopus:105051389114
ISBN
979-8-3315-4413-3
DOI
10.1109/NewCAS64543.2026.11674051
language
English
LU publication?
yes
id
3442ea4d-270f-4e1e-ac4e-46dea5478659
date added to LUP
2026-10-05 13:11:16
date last changed
2026-10-08 13:48:46
@inproceedings{3442ea4d-270f-4e1e-ac4e-46dea5478659,
  abstract     = {{<p>The Posit number system features tapered precision, providing higher accuracy for values near unity while maintaining a wide dynamic range, making the format an attractive alternative to IEEE 754 for machine-learning workloads. This work presents CPAU, a unified and configurable Posit arithmetic architecture that integrates a quire-based MAC unit and a general-purpose execution unit supporting addition, multiplication, division, exponential, and square root. CPAU employs a shared decode-execute-encode pipeline across all operators, substantially reducing hardware cost while preserving accuracy in serial computations. A dedicated exponential unit enables full softmax evaluation directly in the Posit domain without format conversions. Synthesis in a 22 nm technology demonstrates significant area improvements of more than 34% and 50% for multiplier and adder area, respectively. Moreover, the implementation of exponential function provides the functionality required for inference and training.</p>}},
  author       = {{Wu, Yue and Mammadzada, Fuad and Westring, Kristoffer and Rodrigues, Joachim}},
  booktitle    = {{24th IEEE Interregional NEWCAS Conference, NEWCAS 2026 : an IEEE CASS flagship conference. June 21-24, 2026, Saguenay, Québec, Canada:  proceedings}},
  isbn         = {{979-8-3315-4413-3}},
  keywords     = {{Decode; Encode; Exponential; ML training; Posit arithmetic; Softmax}},
  language     = {{eng}},
  publisher    = {{IEEE - Institute of Electrical and Electronics Engineers Inc.}},
  series       = {{24th IEEE Interregional NEWCAS Conference, NEWCAS 2026 - Proceedings}},
  title        = {{CPAU : a hardware efficient configurable posit arithmetic unit}},
  url          = {{http://dx.doi.org/10.1109/NewCAS64543.2026.11674051}},
  doi          = {{10.1109/NewCAS64543.2026.11674051}},
  year         = {{2026}},
}