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 Tools annual

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 Tools monthly

Global: Zendy Plus monthly

Presents the access of premium content as premium feature

Premium Content

Global: Zendy Plus annual

Global: Zendy Free Plan

Precision scaling has emerged as a popular technique to optimize the computeand storage requirements of Deep Neural Networks (DNNs). Efforts towardcreating ultra-low-precision (sub-8-bit) DNNs suggest that the minimumprecision required to achieve a given network-level accuracy variesconsiderably across networks, and even across layers within a network,requiring support for variable precision in DNN hardware. Previous proposalssuch as bit-serial hardware incur high overheads, significantly diminishing thebenefits of lower precision. To efficiently support precisionre-configurability in DNN accelerators, we introduce an approximate computingmethod wherein DNN computations are performed block-wise (a block is a group ofbits) and re-configurability is supported at the granularity of blocks. Resultsof block-wise computations are composed in an approximate manner to enableefficient re-configurability. We design a DNN accelerator that embodiesapproximate blocked computation and propose a method to determine a suitableapproximation configuration for a given DNN. By varying the approximationconfigurations across DNNs, we achieve 1.17x-1.73x and 1.02x-2.04x improvementin system energy and performance respectively, over an 8-bit fixed-point (FxP8)baseline, with negligible loss in classification accuracy. Further, by varyingthe approximation configurations across layers and data-structures within DNNs,we achieve 1.25x-2.42x and 1.07x-2.95x improvement in system energy andperformance respectively, with negligible accuracy loss.

Ax-BxP: Approximate Blocked Computation for Precision-Reconfigurable Deep Neural Network Acceleration