Can LLMs design hardware? | | | | Turtles AI

Can LLMs design hardware?
DukeRem
  Researchers tap ChatGPT-4 to generate functional #Verilog #code, but find #LLMs require #guidance, #feedback & #hardware-specific #training to reliably create complete designs #AI #HardwareDesign #HDL #ConversationalAI. Verilog is a hardware description language used to model electronic systems. Typically, designers use Verilog to describe the digital circuits that make up a system and to simulate and verify the design before manufacturing. The AI-written Verilog experiments detailed in the scientific paper demonstrate enormous potential as well as clear limitations for using ChatGPT -4 and other LLMs in hardware design processes going forward. On the positive side, when ChatGPT-4 had clear and detailed human feedback during the design phase, it was able to rapidly generate functional Verilog code for complex structures like the 8-bit microprocessor mentioned. Where it struggled though was in creating testbench and verification code without additional training data - pointing to a need for hardware-specific conversational AI models. The researchers recommend using ChatGPT-4 as an "effort multiplier" by having it generate initial design drafts that engineers then refine and iterate on - an approach they argue freed up designer time during the implementation of common modules. Overall, the ability to go from specification to functional design quickly with ChatGPT-4's assistance could accelerate prototyping and experimentation. However, the authors stress that the AI models were not yet capable of designing hardware solely from verification tool feedback, implying current LLMs require more targeted training to reliably correct design errors autonomously. Looking ahead, future work investigating LLMs potential at a larger scale with performance metrics could provide further insights to inform the development of specialized hardware-focused conversational AI. Highlights:
  • ChatGPT-4 generated functionally correct Verilog when provided detailed feedback and specifications
  • Testbench and verification code creation fell short, indicating the need for hardware-specific training data
  • Researchers recommend using conversational AI as an "effort multiplier" by refining and iterating on initial AI-generated drafts
  • Current LLMs not yet capable of fully autonomous hardware design from verification tool feedback alone