Session

DescriptionAs AI workloads expand from always-on edge devices to large-scale cloud deployments, achieving performance within strict energy and scalability constraints requires holistic hardware-software co-design. This session highlights innovations enabling adaptive, efficient AI across heterogeneous platforms – from sub-milliwatt always-on vision for wearable devices to adaptive orchestration of RAGs on mobile SoCs to hybrid DRAM–SOT-MRAM processing-in-memory architectures with multi-objective dataflow optimization for LLM inference. At the system scale, learning-based workload scheduling improves cloud energy efficiency, while hardware-software co-design reduces communication overhead in neuromorphic accelerators. Newly learned, degree-aware graph storage structures significantly accelerate dynamic updates and analytics, and distributed parallelization frameworks make graph transformer training scalable across multi-GPU systems. Collectively, these contributions demonstrate the cross-layer innovations required to advance scalable, energy-efficient AI from edge to cloud.