Research on OpenCL optimization for FPGA deep learning application.

Area of Science:

• Computer Science
• Artificial Intelligence
• Hardware Acceleration

Background:

• Deep learning excels in high-dimensional data but faces computational challenges.
• Field-Programmable Gate Arrays (FPGAs) offer energy efficiency and reconfigurability for deep learning.
• Existing OpenCL optimizations for CPUs/GPUs are not directly transferable to FPGAs, hindering performance.

Purpose of the Study:

• To address the performance limitations of implementing deep learning algorithms on FPGAs.
• To propose a novel OpenCL computational model tailored for FPGA architectures.
• To optimize the computationally intensive convolution layer in deep learning models.

Main Methods:

• Development of an OpenCL computational model utilizing an FPGA template architecture.
• Implementation of deep learning algorithms, specifically the convolution layer, using the proposed model.
• Comparative performance analysis against existing optimization programs from Xilinx.

Main Results:

• The proposed OpenCL computational model significantly accelerates deep learning tasks on FPGAs.
• Performance gains ranged from 8 to 40 times compared to Xilinx's optimization programs.
• Demonstrated the effectiveness of the FPGA-tailored model for convolution layer optimization.

Conclusions:

• The developed OpenCL computational model provides a viable solution for high-performance deep learning on FPGAs.
• This approach overcomes the limitations of direct OpenCL tool and model application on FPGAs.
• The model offers a practical method for software programmers to achieve rewarding performance in FPGA-based deep learning implementation.