Presentation
TIDE: A Triangular Interaction-Symmetric Dual-Systolic Engine for Efficient Acceleration of Physics-Inspired Optimization
DescriptionPhysics-inspired optimization (PIO) provides a unified framework for modeling structured interactions inspired by principles from statistical physics, with applications spanning machine learning, operations research, and VLSI design. Representative models—including Hopfield neural networks (HNNs), Boltzmann machines (BMs), and Ising machines (IMs)—encode objectives as energy landscapes shaped by symmetric pairwise couplings. However, conventional hardware accelerators for these applications typically rely on dense matrix-multiplication datapaths that fail to exploit such symmetries, resulting in redundant multiply-accumulate (MAC) operations, excessive weight storage, and input staging delays. To address these inefficiencies, we propose TIDE, a triangular interaction-symmetric dual-systolic engine that structurally exploits symmetry and sparsity in PIO workloads. TIDE stores only the upper-triangular portion of the coupling matrix and performs a two-pass systolic traversal that aligns all data movements in both time and space, eliminating the input staging delays that conventional systolic arrays address using synchronization first-in-first-out (FIFO) buffers. The architecture supports both binary and ternary input representations and incorporates an optional row-bypass path with compact 2-bit encoding to reduce datapath complexity and storage overhead without modifying the array structure. Compared to conventional weight-stationary (WS) and output-stationary (OS) architectures, TIDE achieves 33% lower latency and 50% higher throughput than WS, and nearly 2× throughput over OS. At problem size N=2000, it reduces register storage by 60.0% vs. WS (8-bit), 62.6% vs. OS (8-bit), and 50.0% vs. DiP. With 2-bit encoding, TIDE further provides a 4× storage reduction, yielding 90.0% savings vs. WS (8-bit), 90.7% vs. OS (8-bit), and 87.5% vs. DiP (8-bit).
Event Type
Work in Progress
TimeMonday, July 276:50pm - 6:50pm PDT
LocationExhibit Hall
Similar Presentations