DeepDive

Algorithm/Architecture Co-Design for FPGA-Based DNN Acceleration

DeepDive is an integrative algorithm/architecture co-design framework for deep separable convolutional neural networks on FPGAs.

Key Features

  • Fully functional framework for agile power-efficient execution of DSCNNs
  • Synthesis targeting Xilinx ZCU102 FPGA
  • MobileNetV2 acceleration with up to 8.6x FPS/W improvement
  • Model optimization: quantization (4-bit), layer fusion, pruning, activation approximation
  • Hardware optimization: pipelining, window buffering

Publication

M. Baharani, U. Sunil, K. Manohar, S. Furgurson, and H. Tabkhi, “DeepDive: An Integrative Algorithm/Architecture Co-Design for Deep Separable Convolutional Neural Networks,” GLSVLSI 2021.

Paper Code