..

Summary of LLM's verilog knowledge

  • Testing the “knowledge” of LLMs on verilog
  • Tested on three things - Knowledge, parsing, generation
  • Claude seems better on knowledge and parsing and even simpler generation
  • GPT5 and Claude seems to be the same on bigger designs.

Inspection of mistakes

ALU

  • For the design of an ALU both claude and GPT5 made the same mistake on the LUI instruction. They both used the second operand despite given clear instruction to use the first.