A systematic review of machine learning-driven design space exploration in high-level synthesis

Abstract

In today’s rapidly evolving technological landscape, digitization across all sectors has intensified the complexity of hardware design processes due to increasing demands in data processing, artificial intelligence integration, and performance requirements. Innovative technologies such as the Internet of Things (IoT), 5G networks, big data analytics, and cloud computing introduce multifaceted requirements related to performance, flexibility, adaptability, and energy efficiency. These advancements present significant challenges in digital system design, requiring sophisticated solutions beyond traditional approaches. Conventional design methodologies show limitations in addressing the growing complexity and the resource-intensive nature of the design process. High-Level Synthesis (HLS) has emerged as a critical technology to accelerate digital system design and to enable the hardware implementation of complex algorithms. HLS converts software-level algorithmic specifications into hardware implementations, offering substantial time and cost benefits. However, the increasing complexity of modern digital systems has revealed the limitations of traditional design space exploration (DSE) methods, particularly in optimizing diverse design parameters. Machine learning-based DSE approaches offer transformative solutions, delivering improved efficiency and optimization capabilities in HLS workflows. This study provides a comprehensive analysis of ML-driven DSE techniques in HLS, focusing on innovative approaches to hardware design optimization. It evaluates the impact of various machine learning paradigms — including supervised learning, deep learning, reinforcement learning, and transfer learning — on optimizing critical metrics such as performance, energy efficiency, and resource utilization. ML-based DSE methods demonstrate high accuracy and computational efficiency across vast, multidimensional design spaces, significantly reducing manual efforts through data-driven decision mechanisms. This facilitates rapid evaluation of design parameters and improves development efficiency. Furthermore, ML-driven predictive models accelerate design workflows while reducing computational overhead from synthesis and simulation, enabling accurate predictions for performance, resource use, and energy consumption. The study also explores ML-based DSE contributions in multi-objective optimization, memory and power efficiency, and hardware accelerator design, emphasizing the role of advanced techniques such as Graph Neural Networks (GNNs) in modeling parameter interactions within HLS workflows. The paper concludes with a thorough discussion of the advantages, existing limitations, and future directions of ML-based DSE methods in HLS, highlighting their potential to enhance HLS workflows and meet the evolving demands of high-performance digital system design.