When an AI accelerator slows your intrusion detector: Raspberry Pi 5 and

A study published Nov. 3, 2026 in Discover Internet of Things examines when offloading inference to a dedicated AI accelerator helps—or hurts—in edge IoT intrusion detection. Researchers led by Yasser Hmimou tested a Raspberry Pi 5 paired with a Hailo-8 accelerator, comparing CPU-only inference against accelerator execution using two intrusion-detection datasets (CICIoT2023 and Edge-IIoTset) and three model families. The results are “asymmetric”: for lightweight MLPs, the CPU outperformed the accelerator in streaming throughput, with CPU scaling far ahead at batch size 256, while HailoRT stayed around 37,000 samples/s. For 1D CNNs, HailoRT reversed the outcome with multi-x throughput advantages, depending on the CPU reference regime.