Accelerating ML via Computational Storage

• Unité de recherche: INNOSUISSE
• Numéro de projet: 129.586 INNO-ICT
• Titre du projet anglais: Accelerating ML via Computational Storage

Mots-clé (Anglais)

ICT, Machine learning

Résumé des résultats (Abstract) (Anglais)

GPUs excel in ML but have limited HBM memory, restricting model sizes. Our devices enhance GPU memory by swiftly paging model weights, freeing up HBM for larger models or extended context in LLMs. We aim to boost our prototype's bandwidth by experimenting with faster and larger memories.