Table of Contents
Bridging the Cross-Stack Gap for Processing-in-Memory, Yiwei Zhao (CMU)
Friday, March 20, 2026
Location: CIC 4105 Panther Hollow
Time: 11:00AM-12:00PM
Zoom: https://cmu.zoom.us/j/92411888665?pwd=uwnUtjN7iKKEv2gUV9cE0cyjvanloa.1
Abstract
In the era of massive parallelism, designing efficient systems for parallel platforms and specialized accelerators is both increasingly important but inherently challenging. Prior architecture and hardware work often adopts algorithms and software systems designed for conventional platforms, which are suboptimal for emerging hardware. Conversely, algorithmic and software proposals often rely on idealized assumptions that do not translate well to real systems. As a result, substantial performance opportunities are often left on the table across the design stack.
To bridge this cross-stack gap, I advocate a hardware-software co-design methodology grounded in principled analysis—leveraging provable guarantees throughout the stack to guide practical system design. The design process will be illustrated through an example of PIM-friendly database indexes—B-trees, radix trees, and space-partitioning structures—that balance communication and utilization in practice while providing theoretically provable guarantees under arbitrary data and query distribution. On a 2,560-module real-world PIM system, these designs achieve up to 59× speedups over prior work. We will further discuss extending these techniques to broader heterogeneous architecture or distributed systems, laying foundations for future cross-stack architecture design.
Bio
Yiwei Zhao is a final-year Ph.D. student in the Department of Electrical and Computer Engineering at Carnegie Mellon University, advised by Prof. Phillip B. Gibbons. His research interests focus on cross-stack co-design for emerging hardware systems such as Processing-in-Memory, and broadly span system and architecture design.