Neural Network Quantization Overview

Neural Network Quantization 101

Table of Contents

1. Introduction
2. Int8 vs Int4
3. Matrix Form

1. Introduction

2. INT8 vs INT4

Quantization Ranges

Ranges	INT8	INT4
Input range (fp)	[-1, 1]	[-1, 1]
Weight range (fp)	[-0.5, 0.5]	[-0.5, 0.5]
Quant range	[-128,127]	[-8,7]

Quantization

Steps	INT8	INT4
$s_x$	2/255 ≈ 0.007843	2/15 ≈ 0.1333
$s_w$	1/255 ≈ 0.0039216	1/15 ≈ 0.0667
$q_x$	96	6
$q_w$	102	6
$q_y$	9792	36
$y = s_x * s_w * q_y$	0.3006	0.3200 p
Error	+0.0006	+0.0200

3. Matrix Form