Automated Design Space Exploration and Generation of AI Accelerators

Learn about automated design space exploration and generation of AI accelerators in this conference talk from the Simons Institute's Algorithmic Foundations for Emerging Computing Technologies Boot Camp. Discover how high-level synthesis (HLS)-based frameworks can address the scalability challenges of designing ad hoc accelerators for rapidly evolving machine learning models. Explore how architectural parameters such as datatype, scaling granularity, compute parallelism, and memory sizes can be used to generate performant, fabrication-ready accelerators. Examine real-world applications through accelerators that have been successfully taped out in multiple chips targeting various workloads including convolutional neural networks and transformer networks. Understand how this generator framework can also serve as a benchmarking tool for designs that leverage emerging technologies, providing insights into the future of high-performance and energy-efficient AI accelerator design.