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

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

Explore cross-architecture dynamic migration challenges in heterogeneous acceleration systems through this informative conference talk. Learn about China Mobile's innovative practices for enabling smooth AI application migration across diverse accelerators like GPUs, FPGAs, and NPUs. Discover solutions to tool chain and vendor lock-in issues, including a unified abstraction for accelerators, a middle-compiler approach for cross-architecture compilation, and a runtime supporting dynamic and replaceable linking. Gain insights into the architecture design, open source plans, and a practical demonstration showcasing how applications can migrate freely between accelerators without altering development habits.