• Figure 1: Unified hybrid architecture showing parallel temporal (LSTM-SNN) and spatial (CNN-QNN) processing paths with dynamic fusion. The feature router employs attention mechanisms to optimally allocate processing based on signal characteristics. Here we propose a new approach to operate spiking neural networks (SNN) and quantized neural networks (QNN) together on RISC-V platforms, a promising advancement for medical and industrial IoT applications. Our work is based on three key innovations: dedicated hardware extensions (RV32X-SQ) to jointly accelerate SNN and QNN, compression techniques that reduce the memory footprint by up to 9.1 KB (a saving of 62%), and a predictive thermal management system that maintains a temperature difference (Δ T ) below 5°C for medical and 7°C for industrial. Our dual-channel architecture processes data in parallel, combining temporal analysis via an LSTM-SNN and spatial analysis via a quantized CNN-QNN, all dynamically merged. The results are convincing: on the MIT-BIH arrhythmia data base, we achieve an accuracy of 97.59% with a latency of 8.2 ms and an estimated consumption of only 21 mW, i.e. an efficiency 3.2 times greater than a solution based on ARM Cortex-M7. For industrial bearing fault detection, the accuracy is 97.8%. This solution, compliant with AAMI EC57:2025 and IEC 60601-2-47:2024 standards, sets new benchmarks in terms of efficiency for embedded artificial intelligence.
Lengui et al. (Sun,) studied this question.