Nvidia has shared the development of a large language model (LLM) that can serve as an aid for semiconductor engineers. The tool is expected to enhance the intricate process involved in the design and development of future chips.
Nvidia's announcement comes in the wake of AI increasingly finding its way into chip design processes. Companies like Google have been employing machine learning to enhance its Tensor Processing Unit (TPU) accelerator family. Similarly, chip design software giants like Synopsis and Cadence have reportedly been integrating AI into their applications. Nvidia has especially been instrumental in the sphere of GPU-accelerated lithography tooling.
The recently released paper by Nvidia outlines the role of this generative AI in future chip development. While the AI is not yet released, it is expected to guide future endeavors in building similar chat-based systems or bots.
Evolving AI to Simplify Chip Design
Designing microprocessors involves multiple teams working on different aspects of the chip blueprint. Nvidia researchers demonstrated the use of the company's NeMo framework to modify a 43 billion parameter foundation model. This was done using data relevant to chip design and development, said to be more than a trillion tokens, representing parts of words and symbols.
The model underwent two refining stages of training. The first stage processed 24 billion tokens worth of internal design data and the second stage used 130,000 conversation and design examples. The resulting ChipNeMo models, consisting of seven billion and thirteen billion parameters, then powered three AI applications, including two similar to ChatGPT and GitHub Copilot. The bots have been developed to generate System Verilog code, answer questions on processor design and testing techniques, write scripts for design process automation, and analyze silicon-level bug reports.
The Future of AI in Advanced Chip Development
The goal behind the development of the generative AI is to demonstrate its potential use beyond writing normal app code, into the realms of Verilog and semiconductor engineering. Nvidia's Chief Scientist, Bill Dally emphasised the potential of LLMs in the complex task of designing semiconductors. Despite leveraging AI, Nvidia also pointed out the prime role of skilled professionals in driving the process of semiconductor design.
Mark Ren, the Nvidia researcher who led the project, believes that AI will gradually play a more significant role in advanced chip development. This marks an important stride in Nvidia's ongoing efforts in utilizing accelerated computing and machine learning for semiconductor development, a concept propagated by CEO Jensen Huang.