Addressing Challenges of Cross-Architecture Dynamic Migration Over Heterogeneous Acceleration Systems

Explore a conference talk addressing the challenges of cross-architecture dynamic migration over heterogeneous acceleration systems. Dive into China Mobile's innovative practices for enabling AI applications to migrate smoothly between diverse accelerators such as GPUs, FPGAs, and NPUs. Learn about a unified abstraction for various accelerators, a middle-compiler approach utilizing existing compilers like CUDA, ROCm, and oneAPI to achieve cross-architecture compilation, and a runtime supporting dynamic and replaceable linking. Discover how these solutions aim to overcome tool chain and vendor lock-in issues, allowing applications to migrate freely between different accelerators without altering development habits. Gain insights into the architecture design, open source plans, and witness a demonstration of this groundbreaking approach to cross-system application development and migration.