Exploring the Potential of Flexible 8-bit Format: Design and Algorithm

TL;DR

This paper investigates the advantages of FP8 neural network quantization, compares it with INT, and introduces a flexible mixed-precision framework that optimizes quantization formats for different architectures, improving efficiency across multiple tasks.

Contribution

It presents a novel flexible mixed-precision quantization framework supporting various number systems, enabling optimal format selection for different neural network architectures.

Findings

FP8 quantization offers competitive performance with full precision.

The proposed framework adapts to various neural network tasks.

Experimental results validate the effectiveness of the mixed-precision approach.

Abstract

Neural network quantization is widely used to reduce model inference complexity in real-world deployments. However, traditional integer quantization suffers from accuracy degradation when adapting to various dynamic ranges. Recent research has focused on a new 8-bit format, FP8, with hardware support for both training and inference of neural networks but lacks guidance for hardware design. In this paper, we analyze the benefits of using FP8 quantization and provide a comprehensive comparison of FP8 with INT quantization. Then we propose a flexible mixed-precision quantization framework that supports various number systems, enabling optimal selection of the most appropriate quantization format for different neural network architectures. Experimental results demonstrate that our proposed framework achieves competitive performance compared to full precision on various tasks, including image classification, object detection, segmentation, and natural language understanding. Our work furnishes critical insights into the tangible benefits and feasibility of employing FP8 quantization, paving the way for heightened neural network efficiency in tangible scenarios. Our code is available in the supplementary material.