Low-Power FPGA Architectures for Edge Inference in Portable Air-Quality Monitors

Abstract

This study benchmarks three quantized convolution pipelines mapped to a mid-range Artix FPGA for on-device particulate matter classification. The student-led team optimized memory movement with line-buffer reuse and demonstrated a 38% reduction in dynamic power compared with baseline HLS-generated kernels while retaining 94.1% classification accuracy. Field tests in school-zone corridors indicate stable latency under variable thermal conditions, supporting practical deployment in battery-constrained sensing stations.

Figure 1 compares three synthesized accelerator designs and highlights the proposed architecture as the most efficient operating point for mobile inference workloads.