Methods for Constructing High-Performance Interconnects in System-on-Chip Designs
Kirill D. Lyubavin, Dmitry Vladimirovich Telpukhov · 2024
This paper explores the development of high-performance interconnects essential for System-on-Chip (SoC) designs, focusing on high bandwidth, low latency, and configuration flexibility. We outline key requirements for implementing these interconnects, emphasizing effective methods to achieve optimal system performance. The study includes an experimental investigation of how slave device base addresses impact the performance of programs on embedded RISC- V cores. Key findings suggest that address alignment significantly influences the processor's efficiency in handling transactions. We propose several techniques to enhance interconnect performance: parallel arbitration mechanisms, node prioritization, pipelining, and overhead minimization. The implementation of weighted round-robin arbitration ensures that high-priority data paths maintain efficiency even under heavy load. The use of pipelining and burst transactions in modern interfaces like AMBA AXI and Avalon-MM dramatically increases throughput by reducing clock cycle intervals between data packets. Overhead minimization strategies include optimizing data bus width conversion and interface protocol translation. Additionally, we demonstrate that aligning subordinate device addresses within specific boundaries can reduce the instruction count for address formation, thus improving overall processing speed. Our findings underscore the importance of optimizing interconnect designs to meet the stringent performance demands of contemporary SoCs, ensuring that embedded processing units can interact efficiently with peripheral devices. These methods are recommended for developing high-performance interconnects that support advanced SoC applications.