Compiler Optimization for PIM

Heelim Choi - Core Lab, EE, Yonsei University

Processing-in-Memory (PIM) architectures have emerged as a promising approach to overcome the memory wall, yet their practical benefits remain difficult to realize due to immature software support and limited optimization scope. Existing approaches focus on individual kernels, often overlooking system-level inefficiencies that arise across execution boundaries.

In this seminar, PICO, a compiler framework for cross-kernel optimization in PIM systems, is presented. By coordinating execution beyond individual kernels, PICO reduces overheads and improves resource utilization under the constraints of PIM hardware. Built on modern compiler infrastructure, it provides an end-to-end pipeline from high-level computations to efficient execution.

Evaluations on representative machine learning workloads demonstrate substantial performance improvements over conventional methods, highlighting the importance of system-level coordination in unlocking the full potential of PIM architectures.

CV

Catering Courtesy of AI-ISL