Large Language Models for Verilog Generation and Verification: A Review of Cloud, Domain-Specialised and Resource-Constrained Approaches
Abstract
Register-transfer-level (RTL) design is a critical but labour-intensive stage in digital hardware development. Large language models (LLMs) have recently shown the ability to convert natural-language specifications into Verilog code, thereby reducing manual design effort. This paper presents a review of LLM-based Verilog generation and verification approaches, organised around deployment requirements and verification strategies. The review covers classical electronic design automation (EDA) foundations, general-purpose cloud coding assistants, Verilog-specialised models and benchmarks, feedback-driven and agentic EDA frameworks, open-source verification back ends, and resource-constrained deployment using quantisation and lightweight inference engines. The literature indicates that cloud- and GPU-based systems remain dominant, while compiler and simulator feedback can improve the syntactic validity of generated designs. However, functional correctness remains difficult because compilation alone does not verify intended hardware behaviour. Resource-constrained deployment is feasible in selected cases; for example, RTLCoder reports a 4-bit quantised model of approximately 4 GB that can operate on a laptop. Nevertheless, consistent reporting of latency, memory use, energy consumption, offline operation, privacy, and reproducibility is still limited. This review identifies major open challenges, including functional verification, benchmark quality, scarcity of high-quality Verilog training data, security of generated hardware, small-model efficiency, and standardised resource-aware evaluation. The findings suggest that reliable AI-assisted RTL design will require closer integration of LLMs with deterministic EDA tools, stronger verification oracles, improved benchmarks, and transparent evaluation practices.