Mixed-Signal In-Memory Multi-bit Matrix-Vector Multiplication - Archive ouverte HAL Accéder directement au contenu
Communication Dans Un Congrès Année : 2021

Mixed-Signal In-Memory Multi-bit Matrix-Vector Multiplication

Résumé

The applications for artificial intelligence are wide and cover multiple domains including industry, health, home automation, consumer electronics, automotive, and smart cities. Application-specific integrated circuits performing tiny machine learning at ultra-low power and high accuracy are needed. The Von Neumann wall forces us to shift the processing elements closer to the memory to prevent data movement and therefore reduce energy consumption. Matrix-Vector Multiplication (MVM) can be achieved with many approaches that perform well with binary weights but not with multi-bit multiplications. This paper tries to highlight the advantages of using current sources to perform in-memory computing, improving further the energy consumption to perform multi-bit MVM.
Fichier principal
Vignette du fichier
Mixed-Signal In-Memory Multi-bit Matrix Vector Multiplication - Kevin Herisse.pdf (222.63 Ko) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-03515025 , version 1 (06-01-2022)

Identifiants

  • HAL Id : hal-03515025 , version 1

Citer

Kévin Hérissé, Benoit Larras, Bruno Stefanelli, Antoine Frappé, Andreas Kaiser. Mixed-Signal In-Memory Multi-bit Matrix-Vector Multiplication. 15ème Colloque National du GDR SOC2, Jun 2021, Rennes, France. ⟨hal-03515025⟩
60 Consultations
25 Téléchargements

Partager

Gmail Facebook X LinkedIn More