SA-Kura: An Energy-Efficient Systolic Array Accelerator for Locally-Coupled Kuramoto Drift in Diffusion Sampling
Open Access
- 5 August 2026
- conference paper
- Published by Association for Computing Machinery (ACM)
- p. 1-7
Abstract
Diffusion inference remains costly for edge deployment, yet existing accelerators focus almost exclusively on score networks because standard drift is merely a trivial linear scaling. Kuramoto orientation diffusion replaces this drift with locally coupled phase interactions, improving sampling efficiency but introducing a new hardware bottleneck: a center-dependent nonlinear 5 × 5 stencil evaluated at every reverse step that maps poorly to conventional CNN accelerators and matrix-oriented engines. We present SA-Kura, to our knowledge the first digital systolic-array accelerator dedicated to locally coupled Kuramoto drift. By reformulating pairwise sinusoidal coupling into center-independent neighbor accumulation followed by a single center-dependent multiply-subtract, SA-Kura eliminates in-PE transcendental units and enables register-level-reuse systolic execution. Implemented in synthesizable RTL, integrated into a lightweight RISC-V-based SoC, prototyped on FPGA, and evaluated through 45 nm CMOS synthesis and power analysis, SA-Kura reduces drift-kernel latency and energy by 193× and 69.4× over same-SoC processor software, and is 6.57× faster with 46.0× lower energy per pixel than a Jetson Orin Nano CUDA implementation.Keywords
All Related Versions
Funding Information
- National Research Foundation of Korea (NRF) (RS-2024-00345668)
This publication has 8 references indexed in Scilit:
- EdgeDiff: Energy-Efficient Multi-Modal Few-Step Diffusion Model Accelerator Using Mixed-Precision and Reordered Group QuantizationIEEE Journal of Solid-State Circuits, 2025
- E-BTS: A low-power Event-driven Blink Tracking System with hardware-software co-optimized design for real-time driver drowsiness detectionAlexandria Engineering Journal, 2025
- Efficient Diffusion Models: A Comprehensive Survey From Principles to PracticesIEEE Transactions on Pattern Analysis and Machine Intelligence, 2025
- Designing Low-Power RISC-V Multicore Processors With a Shared Lightweight Floating Point Unit for IoT EndnodesIEEE Transactions on Circuits and Systems I: Regular Papers, 2024
- A mixed-signal oscillatory neural network for scalable analog computations in phase domainNeuromorphic Computing and Engineering, 2023
- A 1,968-node coupled ring oscillator circuit for combinatorial optimization problem solvingNature Electronics, 2022
- Developing TEI-Aware Ultralow-Power SoC Platforms for IoT End NodesIEEE Internet of Things Journal, 2020
- TEI-NoC: Optimizing Ultralow Power NoCs Exploiting the Temperature Effect InversionIEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 2017