Optimization of Deep Neural Networks Using SoCs with OpenCL.

Area of Science:

• Computer Science
• Artificial Intelligence
• Hardware Acceleration

Background:

• Optimizing deep neural networks (DNNs) with evolutionary algorithms (EAs) faces efficiency-vs-flexibility trade-offs.
• Software solutions on general-purpose processors are slow; hardware solutions on heterogeneous platforms (CPUs, GPUs, FPGAs) use diverse methods and languages.
• Existing heterogeneous implementations often lack unified methodologies and criteria.

Purpose of the Study:

• To demonstrate the necessity of a heterogeneous (CPU-GPU-FPGA) platform for accelerating artificial neural network (ANN) optimization via genetic algorithms.
• To present OpenCL-based implementations for evaluating individuals in genetic algorithms on CPU and FPGA platforms.
• To evaluate the efficiency of remote, low-cost FPGA systems-on-a-chip (SoCs) for this optimization task.

Main Methods:

• Utilized a heterogeneous computing approach combining CPUs, GPUs, and FPGAs.
• Developed implementations using OpenCL for calculations related to individuals in genetic algorithms.
• Deployed implementations on both CPU-based and FPGA-based platforms, including remote, low-cost FPGA SoCs.

Main Results:

• Confirmed the need for heterogeneous platforms to accelerate ANN optimization with genetic algorithms.
• Achieved comparable implementation logic for individuals across different platforms (CPU and FPGA) using OpenCL.
• Demonstrated good performance-per-watt efficiency with FPGA SoC implementations.

Conclusions:

• Heterogeneous platforms are crucial for efficient evolutionary optimization of ANNs.
• OpenCL provides a viable framework for implementing genetic algorithm calculations on diverse hardware.
• Low-cost FPGA SoCs present a promising solution for energy-efficient ANN optimization.