Session

DescriptionAs Moore’s Law faces the challenges of extreme technology scaling and increasing model complexity, the burden of system reliability is shifting from hardware to intelligent, full-stack orchestration. This session explores the cutting edge of resilient and predictable computing, spanning deep learning accelerators, cloud microservices, and time-sensitive networks. We navigate the 'Outlier Dichotomy' in LLMs and soft errors in NPUs, introducing micro-architectural safeguards that bridge the gap between safety-critical requirements and performance. Beyond reliability, we tackle the 'black box' of GPU warp scheduling and the interference dynamics of SMT-enabled cloud servers. From adaptive ECC for SSDs to thermal-aware task scaling on Edge TPUs, these papers present a unified vision: achieving microsecond-level precision and robust fault tolerance without sacrificing the efficiency demanded by modern AI workloads.