Domain-specific applications, particularly in artificial intelligence (AI), have significantly transformed the computational paradigm, shifting from traditional CPUs to GPUs and specialized accelerators in recent years. Dedicated AI accelerators, in particular, have demonstrated orders-of-magnitude efficiency improvements compared to CPUs and GPUs. However, as AI algorithms evolve rapidly in diversity and scale, these specialized accelerators are increasingly unable to meet the requirements of reusability across a wide range of emerging algorithms. This has resulted in a notable gap between the accelerating pace of AI algorithmic innovation and the adaptability of hardware accelerators. This seminar explores an agile approach to designing programmable and reusable accelerator generators, aiming to strike a balance between efficiency and flexibility in AI acceleration. In addition, the seminar delves into opportunities for co-exploration of hardware architecture, workload scheduling, and code generation, enabling a comprehensive and full-stack agile AI deployment framework.
14/2/2025 11:00 - 12:00
ESAT Aula R
