Full system-on-chip (SoC) design and simulation that incorporates a deep learning accelerator from NVIDIA

Abstract

Deep learning applications are increasingly being deployed on resource-constrained devices, where power efficiency and computational performance are critical. To address this, hardware accelerators such as Nvidia Deep Learning Accelerator (NVDLA) have been developed to provide efficient inference capabilities. This project focuses on the integration of NVDLA into PULPissimo, an open-source low-power System-on-Chip (SoC) platform, using the AXI interface. A custom AXI IP was developed and successfully integrated as a proof of concept prior to incorporating NVDLA. The integration process involved adapting bus connections, configuring address mappings, and building a wrapper module for seamless communication between NVDLA and the SoC. Simulation-based testing was conducted using pretrained LeNet and AlexNet models to evaluate inference functionality. The results demonstrated that NVDLA was able to process both models correctly, with all predictions falling within the top five classifications. Although simulation runtimes were long, the results validate the feasibility of NVDLA integration. Future work includes deploying the integrated system onto an FPGA board for real-time hardware verification, performance benchmarking, and comparative analysis with other NPUs and accelerators.