Deep Learning AMR Model Inference Acceleration with CFU for Edge Systems

Edge computing could benefit many artificial intelligence applications compared to centralized solutions. However, achieving acceptable performance within typical edge devices' computational power limitations is challenging. This research investigates a promising approach to enhance the processing capabilities of edge devices by using the RISC-V Custom Function Unit (CFU) extension to offload computation to dedicated accelerator hardware. The approach was tested on the Automatic Modulation Recognition problem. To solve the problem, we designed a simple Convolutional Neural Network (CNN), which classifies ten different modulations with an accuracy of ~91% at signal-to-noise ratios greater than 0. Then, we developed and synthesized on an FPGA a custom CFU for the RISC-V processor to accelerate the convolutional layer, the most computationally demanding part of the CNN. The inference speedup gained is approximately 144 times - two orders of magnitude, and the inference wall time on the development board is 65 ms. This opens many new applications for such CNNs. Similar approaches to accelerators would be beneficial for other types of neural nets.