English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

Modern deep neural network (DNN) models generally require a huge amount ofweight and activation values to achieve good inference outcomes. Those datainevitably demand a massive off-chip memory capacity/bandwidth, and thesituation gets even worse if they are represented in high-precisionfloating-point formats. Effort has been made for representing those data indifferent 8-bit floating-point formats, nevertheless, a notable accuracy lossis still unavoidable. In this paper we introduce an extremely flexible 8-bitfloating-point (FFP8) format whose defining factors - the bit width ofexponent/fraction field, the exponent bias, and even the presence of the signbit - are all configurable. We also present a methodology to properly determinethose factors so that the accuracy of model inference can be maximized. Thefoundation of this methodology is based on a key observation - both the maximummagnitude and the value distribution are quite dissimilar between weights andactivations in most DNN models. Experimental results demonstrate that theproposed FFP8 format achieves an extremely low accuracy loss of $0.1\%\sim0.3\%$ for several representative image classification models even without theneed of model retraining. Besides, it is easy to turn a classicalfloating-point processing unit into an FFP8-compliant one, and the extrahardware cost is minor.

All-You-Can-Fit 8-Bit Flexible Floating-Point Format for Accurate and Memory-Efficient Inference of Deep Neural Networks