2026 ASIAN Conference on Communication and Networks (ASIANComNet 2026)

This paper presents a portable RTL-based INT8 CNN accelerator for cross-vendor FPGA Edge AI inference. The proposed architecture separates the CNN computational core from platform-dependent components, enabling the same RTL core to be reused across FPGA devices from different vendors. A quantized LeNet-5 network is implemented on the Gowin Tang Nano 20K and Intel DE10-Lite platforms and compared with software-based inference on the ESP32-S3 and ESP32 P4. Experimental results show that both FPGA implementa tions achieve 97.72% classification accuracy on the deployment dataset. The CNN computational core achieves 100% reuse, while the overall Gowin-to-Intel RTL portability reaches 89.45%. At 81 MHz, the Tang Nano 20K and DE10-Lite achieve inference latencies of 4.905 ms and 4.910 ms, corresponding to through puts of 203.9 FPS and 203.7 FPS, respectively. Both FPGA implementations are approximately 3.13 times faster than the ESP32-S3 baseline. These results demonstrate that the proposed architecture enables consistent CNN inference across FPGA devices from different vendors while maintaining a high degree of RTL reuse

Edge AI, CNN accelerator, cross-vendor FPGA, RTL portability, INT8 quantization