Title:  Architecting 3D Intelligence: From Dedicated Accelerators to Heterogeneous Systems through Algorithm-Hardware Co-Evolution

Date: Friday, December 5, 2025

Time: 12:00-14:00 ET

Location:  https://gatech.zoom.us/my/li.sixu 

Sixu Li

Ph.D. Student

School of Computer Science

Georgia Institute of Technology

Committee members

Dr. Yingyan (Celine) Lin, Co-Advisor: College of Computing, Georgia Institute of Technology

Dr. Josiah Hester, Co-Advisor: College of Computing, Georgia Institute of Technology

Dr. Hyesoon Kim: College of Computing, Georgia Institute of Technology

Dr. Tushar Krishna: College of Computing, Georgia Institute of Technology

Dr. Thierry Tambe: Department of Electrical Engineering, Stanford University

Abstract:

The emergence of 3D intelligence systems demands the ability to perceive, reconstruct, and reason about spatial environments in real time and under strict energy constraints. Each stage in this pipeline exhibits a distinct computational profile: perception and reconstruction are dense and regular, rendering combines structured and dynamic access, and reasoning involves highly irregular and memory-intensive computation. This dissertation argues that these algorithmic characteristics naturally correspond to a hardware hierarchy of increasing architectural generality, dedicated accelerators, enhanced GPUs, and heterogeneous GPU–PIM systems, forming a unified framework for 3D intelligence. The dissertation embodies this framework through three concrete systems: Fusion-3D, a domain-specific accelerator that reorganizes neural reconstruction dataflows and spatial tiling for bandwidth-efficient 3D perception. GauRast, an enhanced GPU rasterization pipeline that bridges classical graphics and neural rendering through microarchitectural extensions. ORCHES, a GPU–PIM collaborative system for large-scale reasoning that introduces predictive scheduling and speculative execution to balance computation and memory proximity. Together, these systems establish a co-evolutionary design principle: as algorithmic entropy increases from perception to reasoning, hardware specialization must relax correspondingly, from fixed-function accelerators to flexible heterogeneous systems. This principle not only unifies the presented works but also offers a systematic methodology for designing future hardware that scales with the growing diversity and complexity of 3D intelligent workloads.