Cross-Domain Transfer Learning for Validation Acceleration in Heterogeneous Computing Architectures

Abstract

The increasing heterogeneity in modern computing architectures introduces significant complexity in design validation, especially as diverse hardware accelerators proliferate across domains. This paper investigates the application of cross-domain transfer learning (CDTL) to accelerate the validation process of heterogeneous systems by reusing knowledge from similar validation tasks across different architectural domains. We explore how models trained on one domain (e.g., GPU-based systems) can support validation efforts in another (e.g., FPGA-based systems) and identify key enablers, bottlenecks, and optimization strategies. Our findings suggest that CDTL significantly reduces validation time and resource usage, maintaining high accuracy in bug detection. We provide experimental results, discuss challenges, and present a comparative literature review highlighting the promise of CDTL in hardware-software co-design.