Genetic Algorithm Design Exploration for On-Device Training on FPGAs
Genetic Algorithm Design Exploration for On-Device Training on FPGAs
We propose an automated Design Space Exploration (DSE) workflow for generating adaptive and reconfigurable deep learning models on FPGA hardware. The workflow consists of two main components: Offline Design Exploration (ODE) and Online Design Reconfiguration (ODR). ODE applies a multi-objective genetic algorithm to explore CNN-based hardware configurations, optimizing for latency and resource utilization by leveraging intra-layer parallelism. Given a CNN architecture and user-defined constraints, the hardware model is generated automatically. ODR enables runtime hardware adaptability by dynamically selecting between partial or full reconfigurable designs based on application requirements. This flexibility is essential for time-critical, autonomous onboard systems. We demonstrate the proposed workflow on the Xilinx Zynq-7100 FPGA operating at 200 MHz, using CNN models trained on MNIST, SVHN, and CIFAR-10. ODE-generated designs show latency improvements of up to 95 times for MNIST, 71 times for CIFAR-10, and 18 times for SVHN. Resource utilization in DSP slices was improved by up to 44 times for MNIST, 52 times for SVHN, and 24 times for CIFAR-10. The ODR approach achieved trade-offs between accuracy and performance, such as a 0.7 percent accuracy drop for a 13 times speedup and 25 percent power reduction on MNIST, a 2 percent drop for 14 times speedup and 28 percent power savings on SVHN, and a 4 percent drop for 50 times speedup with 32.5 percent power reduction on CIFAR-10.
Alaa Mazouz、Van-Tam Nguyen
微电子学、集成电路计算技术、计算机技术
Alaa Mazouz,Van-Tam Nguyen.Genetic Algorithm Design Exploration for On-Device Training on FPGAs[EB/OL].(2025-04-11)[2025-04-26].https://arxiv.org/abs/2504.08534.点此复制
评论