RTL code generation from natural language inputs

?

The Gen AI is widely explored in every area these days. I came across this research paper on front end chip design using Generative AI framework called ChipGPT.

?Currently I see this ChipGPT as an intern joined in a team, requires training to gain domain knowledge, split bigger task and smaller tasks and provide required skills to perform the job. During this process, intern also requires feedback or finetuning on outcomes. But, once enough training is completed, intern becomes an asset to the team :).

High level synthesis tools are already able to generate RTL code but they can convert only the high-level language inputs such as C++ code. The use of LLMs enables RTL logic generation from natural language input.

?

The authors in this paper named "ChipGPT: How far are we from natural language hardware design" explained how an power, performance and area (PPA) efficient RTL can be generated. This is achieved by adding additional modules Prompt Manager (PM), Output Manager (OM) and search method.? The prompt manager placed before the GPT model helps to improve the program quality and output manager helps to improve the generated RTL and finally search method helps to meet PPA (power, performance, Area) requirements of generated RTL. The tuning used doesn’t change the original LLM.

?

The chipGPT front end EDA frame work assists humans in compiling specifications to logic design. It works with below steps:

?split(spec) = {eg, if ace, func, compose}

rawcode = GP T(PM(if ace, func, compose, f d))

code = OM(rawcode, EM) codelist = codelist ∪ code

optcode = Search(codelist, requirements)

?

1. Specification split method: Translate the specifications into a formal and unambiguous form
2. Specification serialization method in prompt-manager:? To make these irregular split partitions become an available GPT input sequence
3. Output Manager : To improve the quality of the code(raw program) generated by the GPT model. At this stage typically PPA requirements are not met.
4. Search method : This is compatible with the previous stages that ranks GPT-generated programs and identifies the version according to the requirements.

?

In my view, these kind of frameworks are going to assist front end engineers thus reduces RTL development phase. It is very unlikely to write entire RTL based on natural language specifications. What do you think on these front EDA frameworks?? Feel free to comment.

?Disclaimer: The views or opinions expressed in this article are mine alone and do not represent those of my employer

References :?

Kaiyan Chang, Et all, ChipGPT: How far are we from natural language hardware design 2305.14019.pdf (arxiv.org)

?

?

?