Beschleuniger des Neuronalen Faltungsnetzes (cnn) unter Verwendung von Vorausschauender Faltung

Details

Veröffentlichungs-Nr.

EP4764952

Anmeldetag

23. Dezember 2024

Veröffentlichung

24. Juni 2026

Rechtsraum

EP

Abstract

There is provided an accelerator (1) implementing a convolutional neural network, CNN, to determine an output feature map, in response to the receipt of an input feature map. The accelerator (1) comprises at least one serializer (3) to serialize the input feature map into a series of 1D serialized vectors, and a plurality of layersets (2). The serializer (3) broadcasts the serialized vectors to at least one of the layersets, the layersets being interconnected in a pipelined fashion, each layerset corresponding to a pipeline stage. Each layerset (2) comprises filter cores (22) to perform the equivalent computation of a CNN layer. The output of a given interconnected layerset is directly transmitted to the next interconnected layerset, without storage of the given layerset output in an external memory, the last interconnected layerset being configured to provide the output feature map. This enables merging CNN layers.

Anmelder

Firma

Institut Mines Telecom

Land

🇫🇷 Frankreich