An Adaptive Congestion-aware Approximate Communication (ACAC) scheme and implementation for Network-on-Chip systems

Data-intensive applications like machine learning and pattern recognition impose heavy Network-on-Chip (NoC) communication loads, drastically increasing on-chip latency. Leveraging the error-tolerant nature of such applications, this paper proposes the Adaptive Congestion-Aware Approximate Communication (ACAC) scheme to dynamically balance approximation and reliability by detecting congestion distribution through three integrated strategies — source control (adjusting approximation rates at the source node), on-the-way control (refining decisions at intermediate routers), and a combined strategy. Cycle-accurate simulations shows a 6% to 36% saturated injection rate improvement over ABDTR under synthetic traffic and superior comprehensive performance across all strategies in real-application traces, particularly with the combined strategy achieving the best results while reducing additional control traffic by 52% compared to ABDTR. Despite these enhancements, the suggested framework incurs minimal hardware overhead, with only a 7.42% area increase and 7.79% power consumption rise relative to traditional NoC designs, demonstrating its potential as a scalable, application-aware solution for energy-efficient NoC optimization in error-resilient domains.