Abstract
Energy efficiency is one of the biggest challenges in contemporary AI accelerators. In this context, we present GAVINA: a novel accelerator architecture that leverages the concepts of quantization, mixed-precision and aggressive undervolting. GAVINA is highly flexible and dynamically reconfigurable in terms of precision and approximation, and it can achieve an energy efficiency of over 200 TOP/sW, currently one of the highest in state-of-the-art AI accelerators.
Short bio
Jordi Fornt Mas received the B.S. degree in industrial electronics from the Universitat Politècnica de Catalunya (UPC), Barcelona, Spain, in 2019, and the M.S. degree in electrical engineering and information technology from the Swiss Federal Institute of Technology in Zürich (ETH Zürich), Zürich, Switzerland, in 2021. He is currently pursuing the Ph.D. degree with the High-Performance Integrated Circuits (HIPICS) Group, UPC.,In 2020, he worked as a Research Intern with the In-memory Computing Group, IBM Research–Zürich, Zürich. In 2021, he joined the Barcelona Supercomputing Center (BSC), Barcelona, where he also conducts research towards his Ph.D. degree in deep learning accelerator design.
Speakers
Speaker: Jordi Fornt Mas. Research Engineer. Synthesis and Physical design of ICs Group - Computer Sciences, BSC