Modern generative AI solutions rely on ever increasing amounts of data. Matching the performance needs of these applications with corresponding improvements in energy efficiency and cost efficiency is a grand challenge. Heterogeneous computing combining different types of accelerators, is by far the most promising approach to addressing this challenge. Accelerators are hardware devices designed for a specific class of workloads.
Philippe Bonnet, a professor at DIKU and the leader of Samsung Semiconductors’ Global Open-ecoSystem Team (GOST), explores the challenges and opportunities associated with hardware / software co-design for the design of future acceleartors, framing the discussion within a business context.