Design neural network accelerator for real time edge inference
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Deep neural networks have achieved remarkable progress across computer vision and scientific computing, but their increasingly intensive computation and data movement requirements expose significant limitations in conventional CPU–GPU architectures. In convolutional neural networks (CNNs), memory access dominates energy consumption, making them challenging to deploy efficiently on power- and bandwidth-constrained edge platforms. This dissertation presents the design and FPGA implementation of a Row-Stationary (RS) CNN accelerator targeting real-time inference of a lightweight residual CNN (DynNet) for two-dimensional convection–diffusion prediction. The accelerator is implemented on a Xilinx Zynq UltraScale+ XCZU15EG MPSoC and features an 8×8 PE array, multi-banked feature and weight buffers, a sliding-window generator, and a custom instruction-driven control pipeline. The proposed design exploits RS dataflow together with channel grouping, spatial tiling, and fixed-point INT16/INT8 arithmetic to maximize on-chip data reuse and minimize off-chip memory bandwidth. Experimental results show that the accelerator achieves sub-20ms latency per frame (50+FPS) at a power budget of 8–12W, delivering more than 5× energy savings compared with high-performance CPUs and GPUs. Accuracy remains close to the FP32 PyTorch baseline, with INT16 achieving < 0.1% error and INT8 providing a compact, energy-efficient alternative. These results demonstrate the effectiveness of RS dataflow and FPGA-oriented specialization for real-time edge deployment of CNN-based scientific and vision workloads.
Fields of Study & Subjects
Advanced Neural Network Applications
Embedded Systems Design Techniques
Parallel Computing and Optimization Techniques
Publication details
- Journal / source
- DR-NTU (Nanyang Technological University)
- Publication date
- 1 January 2026
- Language
- EN
- Document type
- Thesis