In-memory-computing CNN accelerator employing charge-domain compute

High-dimensional matrix-vector-multiplications (MVM) are the main operations of deep neural networks (DNN). As the size of DNNs increases, data movement becomes a problem and limits their performance. Analog in-memory computing accelerators are one of the most promising solutions to reduce this prob...

ver descrição completa

Detalhes bibliográficos
Autor: Echeverria Olaiz, Unai
Formato: tesis de maestría
Fecha de publicación:2020
País:España
Recursos:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2117/340559
Acesso em linha:https://hdl.handle.net/2117/340559
Access Level:acceso abierto
Palavra-chave:Neural networks (Computer science)
Energy consumption
AI
ML
DL
NN
CNN
SRAM
Xarxes neuronals (Informàtica)
Energia -- Consum
Àrees temàtiques de la UPC::Enginyeria electrònica
Descrição
Resumo:High-dimensional matrix-vector-multiplications (MVM) are the main operations of deep neural networks (DNN). As the size of DNNs increases, data movement becomes a problem and limits their performance. Analog in-memory computing accelerators are one of the most promising solutions to reduce this problem. This project designs an in-memory computing solution that employs charge-domain compute using 22FDX technology. The design, called multiplying bit cell (M-BC), consists of an 8T bit cell and a MOM capacitor. The design is part of an architecture of 8x8 = 64 neuron tiles that performs the filtering operation of up to 3x3x512 input activation (IA). Each neuron tile is composed of 64x64 = 4096 neuron patches. The design achieves energy efficiency of 1170 TOPS/W and throughput of 18876 GOPS.