Scholarlink

With the development of convolutional neural network, its model scale and data volume are getting bigger and bigger, so the convolutional neural network gas pedal has been born. NVDLA as an open-source project from NVIDIA, supports a variety of classical models. However, due to the complexity of the ASIC design flow and the lack of support for the latest operators, its application scenarios are significantly limited. In this paper, we propose a method to map NVDLA onto FPGA, and employ the Tengine toolchain to replace the original software stack provided by the NVDLA. This approach optimizes the performance of the supported LeNet5 and ResNet18 models. Experimental results show that the Tengine toolchain increases inference speed by 151% compared to the NVDLA website’s software toolchain.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Software-Hardware Co-design Platform for Convolutional Neural Network Accelerator with Tengine

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Zhiwei Liu,
  • Shangwen Sui,
  • Wanyu Lu,
  • Bin Yu,
  • Ao Li

摘要

With the development of convolutional neural network, its model scale and data volume are getting bigger and bigger, so the convolutional neural network gas pedal has been born. NVDLA as an open-source project from NVIDIA, supports a variety of classical models. However, due to the complexity of the ASIC design flow and the lack of support for the latest operators, its application scenarios are significantly limited. In this paper, we propose a method to map NVDLA onto FPGA, and employ the Tengine toolchain to replace the original software stack provided by the NVDLA. This approach optimizes the performance of the supported LeNet5 and ResNet18 models. Experimental results show that the Tengine toolchain increases inference speed by 151% compared to the NVDLA website’s software toolchain.