Beyond Verilog: Agents for Emerging HDLs

Large Language Models (LLMs) are transforming the programming language landscape, but their effectiveness is limited for niche Hardware Description Languages (HDLs) due to insufficient training data. This paper introduces HDLAgent, an agentic AI framework that enhances LLMs’ proficiency with underrepresented HDLs.Our agent employs two key mechanisms: HDL description summaries that facilitate knowledge transfer from familiar languages like Verilog, and enhanced grounding through compiler error feedback. We demonstrate that HDLAgent significantly improves code generation across multiple HDLs. For example, PyRTL’s functional correctness rate improves from 0% to 35% with Mixtral 8x7B, and Chisel’s success rate increases from 0% to 59% with GPT-3.5-turbo. For small code samples (under 25 lines), HDLAgent achieves over 90% success across all tested HDLs. Beyond performance metrics, we provide insights on scalability limitations with increasing code complexity and propose modifications to HDL designs that could better accommodate LLM-based generation.