GANDSE: Generative Adversarial Network-based Design Space Exploration for Neural Network Accelerator Design

With the popularity of deep learning, hardware implementation platform learning has received increasing interest. Unlike general purpose devices, e.g., CPU or GPU, where algorithms are executed at software level, neural network accelerators directly execute to achieve higher energy efficiency and performance improvements. However, as evolve frequently, engineering effort cost designing greatly increased. To improve design quality while saving cost, automation for was proposed, space exploration used automatically search optimized accelerator within a space. Nevertheless, complexity brings dimensions As result, previous no longer effective enough find an design. In this work, we propose framework named GANDSE, rethink problem exploration, novel approach based on generative adversarial (GAN) support high-dimension large The experiments show that GANDSE is able more designs in negligible time compared with approaches including multilayer perceptron reinforcement learning.