计算机科学
互连
电子工程
电气工程
工程类
功率(物理)
工作(物理)
钥匙(锁)
度量(数据仓库)
物理
作者
N. Deepika,S. Suganya,M Adinath,R Vijayalakshmi,A Krishna Veni,E Rohan Kumar
出处
期刊:
[Institute of Electrical and Electronics Engineers]
日期:2026-04-09
卷期号:: 524-529
标识
DOI:10.1109/adsssc67751.2026.11582133
摘要
This paper presents a silicon-validated, learningaugmented AXI4 interconnect integrating a hardware-efficient reinforcement learning (RL)-based arbiter for adaptive Quality-of-Service (QoS) in heterogeneous System-on-Chip (SoC) platforms. Unlike prior simulation-only RL arbiters, this work provides a fully synthesizable implementation where the learned policy is compressed into a lookup-table (LUT)-based inference engine, enabling real-time arbitration without floating-point or neural network overhead. The arbitration problem is formulated as a Markov decision process, where the state captures per-master queue occupancy and latency, actions represent dynamic priority assignments, and the reward optimizes latency, fairness, and utilization. The policy is trained offline using traffic traces derived from heterogeneous AI workloads and mapped to compact combinational logic for on-chip deployment. The design is implemented in SystemVerilog and validated on FPGA and through full ASIC flow in $\mathbf{1 6 ~ n m}$ FinFET technology. Postlayout results achieve 1.6 GHz operation and $102.4 \mathrm{~GB} / \mathrm{s}$ peak bandwidth. Compared to round-robin, fixed- priority, weighted round-robin, and AXI-REALM-style schemes, the proposed arbiter reduces 99th-percentile latency by 55\%, improves throughput by 30%, and sustains 94% utilization. The RL logic incurs $\lt0.02 \mathrm{~mm}^{2}$ area and 14 mW power with no timing impact.
科研通智能强力驱动
Strongly Powered by AbleSci AI