Domain-specific accelerators based on the Explicit Decoupled Data Orchestration (EDDO) paradigm with software-managed scratch-pads are becoming more popular due to their energy efficiency comparedto traditional architectures for workloads that can be analyzed at compile time. By separating data movement from computation and using specialized programmable cores, EDDO-based accelerators, like the IBM AIU [19], enhance performance and energy efficiency across various applications. However, adopting these architectures presents certain chal- lenges, especially for software frameworks. These challenges range from programming languages and compiler optimizations to efficient code generation, requiring fresh solutions from the research community. In this symposium, I will discuss the challenges related to programming and code optimization for EDDO architectures, and present one of the code optimization approaches developed for IBM AIU accelerators, which are designed based on EDDO principles.

Prasanth Chatarasi is a Research Staff Member at IBM T.J. Watson Research Center in Yorktown Heights. He joined IBM in 2020 after completing his Ph.D. at the Georgia Institute of Technology. His current research interests are on compilers and micro-architecture of the next-generation AI hardware technologies. He is currently working on systematic, principled, and modular code generation efforts for the current and future IBM’s AI Accelerators.