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

We show that selecting a single data type (precision) for all values in DeepNeural Networks, even if that data type is different per layer, amounts toworst case design. Much shorter data types can be used if we target the commoncase by adjusting the precision at a much finer granularity. We propose DynamicPrecision Reduction (DPRed), where we group weights and activations and encodethem using a precision specific to each group. The per group precisions areselected statically for the weights and dynamically by hardware for theactivations. We exploit these precisions to reduce: 1) off-chip storage andoff- and on-chip communication, and 2) execution time. DPRed compressionreduces off-chip traffic to nearly 35% and 33% on average compared to nocompression respectively for 16b and 8b models. This makes it possible tosustain higher performance for a given off-chip memory interface while alsoboosting energy efficiency. We also demonstrate designs where the time requiredto process each group of activations and/or weights scales proportionally tothe precision they use for convolutional and fully-connected layers. Thisimproves execution time and energy efficiency for both dense and sparsenetworks. We show the techniques work with 8-bit networks, where 1.82x and2.81x speedups are achieved for two different hardware variants that takeadvantage of dynamic precision variability.

DPRed: Making Typical Activation and Weight Values Matter In Deep Learning Computing