Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

Exploration of a Dynamic Approximate Multiplier for Mixed-Precision Inference

Weidemann, Eivind Aksel LU (2026) EITM01 20252
Department of Electrical and Information Technology
Abstract
The increasing demand for efficient neural network (NN) inference on edge devices has driven the need for hardware-level optimizations that balance computational accuracy with energy and area efficiency. This thesis explores the design and implementation of a dynamic approximate multiplier capable of mixed-precision inference, focusing on floating-point (FP) formats. Using a logarithmic-approximation approach, we implement three precision modes—High Precision Correction (HPC), Low Precision Correction (LPC), and No Correction (NC)—and evaluate their performance on Artix-7 FPGA hardware.

Our results indicate that at 8-bit widths, the numerical error difference between full precision and approximate modes gives a manageable drop off in NN... (More)
The increasing demand for efficient neural network (NN) inference on edge devices has driven the need for hardware-level optimizations that balance computational accuracy with energy and area efficiency. This thesis explores the design and implementation of a dynamic approximate multiplier capable of mixed-precision inference, focusing on floating-point (FP) formats. Using a logarithmic-approximation approach, we implement three precision modes—High Precision Correction (HPC), Low Precision Correction (LPC), and No Correction (NC)—and evaluate their performance on Artix-7 FPGA hardware.

Our results indicate that at 8-bit widths, the numerical error difference between full precision and approximate modes gives a manageable drop off in NN image classification accuracy. Hardware synthesis shows that 8-bit NC and LPC multipliers provide significant advantages in area (number of LUTs) and latency compared to commonly used 8-bit fixed-point multipliers. Furthermore, we demonstrate that a mixed-precision strategy, guided by layer-wise sensitivity, allows for maximizing hardware efficiency while maintaining nearly baseline accuracy. We conclude that approximate 8-bit floating-point multipliers are in some cases a viable alternative to fixed-point arithmetic. (Less)
Popular Abstract
Artificial Intelligence is everywhere, but the "brains" (Neural Networks) behind it are power-hungry. To make AI run faster and with better battery life on your phone or small devices, a bit of mathematical perfection can be sacrificed. This is called Approximate Computing.

This thesis takes a look at how one of the most common tasks in AI computing can be simplified: multiplication. Normally, computers try to be 100\% accurate. However, AI is surprisingly resilient; it can still "recognize" a cat even if the internal math is somewhat off. In this work, a multiplier is designed that can switch between different levels of "accuracy" on the fly.

By using "smaller" numbers in a physical sense and adjustable approximate math, a design is... (More)
Artificial Intelligence is everywhere, but the "brains" (Neural Networks) behind it are power-hungry. To make AI run faster and with better battery life on your phone or small devices, a bit of mathematical perfection can be sacrificed. This is called Approximate Computing.

This thesis takes a look at how one of the most common tasks in AI computing can be simplified: multiplication. Normally, computers try to be 100\% accurate. However, AI is surprisingly resilient; it can still "recognize" a cat even if the internal math is somewhat off. In this work, a multiplier is designed that can switch between different levels of "accuracy" on the fly.

By using "smaller" numbers in a physical sense and adjustable approximate math, a design is implemented that is cheaper and less power-hungry than other alternatives. (Less)
Please use this url to cite or link to this publication:
author
Weidemann, Eivind Aksel LU
supervisor
organization
course
EITM01 20252
year
type
H2 - Master's Degree (Two Years)
subject
keywords
approximate computing, mixed-precision inference, floating-point multiplier, quantization, neural network inference, logarithmic approximation
report number
LU/LTH-EIT 2026-1111
language
English
id
9224789
date added to LUP
2026-03-26 10:28:52
date last changed
2026-03-26 10:28:52
@misc{9224789,
  abstract     = {{The increasing demand for efficient neural network (NN) inference on edge devices has driven the need for hardware-level optimizations that balance computational accuracy with energy and area efficiency. This thesis explores the design and implementation of a dynamic approximate multiplier capable of mixed-precision inference, focusing on floating-point (FP) formats. Using a logarithmic-approximation approach, we implement three precision modes—High Precision Correction (HPC), Low Precision Correction (LPC), and No Correction (NC)—and evaluate their performance on Artix-7 FPGA hardware.

Our results indicate that at 8-bit widths, the numerical error difference between full precision and approximate modes gives a manageable drop off in NN image classification accuracy. Hardware synthesis shows that 8-bit NC and LPC multipliers provide significant advantages in area (number of LUTs) and latency compared to commonly used 8-bit fixed-point multipliers. Furthermore, we demonstrate that a mixed-precision strategy, guided by layer-wise sensitivity, allows for maximizing hardware efficiency while maintaining nearly baseline accuracy. We conclude that approximate 8-bit floating-point multipliers are in some cases a viable alternative to fixed-point arithmetic.}},
  author       = {{Weidemann, Eivind Aksel}},
  language     = {{eng}},
  note         = {{Student Paper}},
  title        = {{Exploration of a Dynamic Approximate Multiplier for Mixed-Precision Inference}},
  year         = {{2026}},
}