[Introduction to New Wisdom]The birth of generative AI has opened up another path for chip design. Now, both technology companies such as NVIDIA and academia are trying to develop AI systems that can design chips completely independently.
Using production-based AI to accelerate chip design will become a cornerstone of the semiconductor industry.
Advertisement
In the past year, computing giant NVIDIA, chip design companies Synopsys, Cadence Design Systems, and academic developers have made many attempts.
They each developed an AI tool——
It is designed to speed up the work of engineers by automating hardware code and verification code, and to help large design teams work together by summarizing notes and status updates.
Letting AI participate in chip design is all due to the artificial intelligence boom in 2023, and the supply of dedicated AI chips has been in a tight state.
At the same time, the end of Moore's Law, which states that the number of transistors in a chip will double approximately every two years, has prompted many companies to explore new chip architectures to produce more specialized chips.
Advertisement
Experts say the U.S. does not have enough engineers capable of designing these advanced chips for AI and specific applications such as self-driving cars and drones, all of which are experiencing growing demand.
NVIDIA ChipNeMo, dedicated to AI chip design
Bryan Catanzaro, vice president of applied deep learning research at Nvidia, said,
Because a GPU can handle thousands of tasks simultaneously, it requires nearly a thousand people to make it, and everyone must understand how the various parts of the design work together while constantly improving.
In this regard,The NVIDIA team developed a new customized large model ChipNeMocapable of performing tasks such as answering questions about GPU architecture, or generating chip design language code.
The researchers trained this AI system based on the open source Llama 2 model. At the same time, the AI system is also designed to work with existing design automation tools such as Synopsys.
Nvidia's internal engineers have been using ChipNeMo for a year, and Catanzaro said they have found the system useful in training junior engineers and summarizing notes and status updates from 100 different teams.
Google, chip design AI companies enter the battle
For Google DeepMind, they also developed an AI system to improve logic synthesis.
This is the stage of chip design that involves translating a description of circuit behavior into an actual circuit. Google said these technologies may be used to improve its own custom artificial intelligence chips, known as “Tensor Processing Units” (TPUs).
In addition, chip design company Synopsys released an AI tool last year-named Synopsys.ai Copilot. This is a tool developed in partnership with Microsoft through OpenAI's large model to help engineers collaborate.
Microsoft's internal silicon teams are using the tool to support their engineering needs, the company said. This AI tool can answer questions about how to use the company's design tools and create workflow scripts.
It can also generate RTL (a chip design language used to standardize chip architecture) with just a conversation in plain English.
Research explosion in academia
In academia, there are also many studies in this direction. Research at several universities, including New York University, is dedicated to discovering other ways to identify generative AI-accelerated chip designs.
Some of the research has been funded by companies such as Synopsys and chip giant Qualcomm.
A team at NYU Tandon School of Engineering designed a chip in one month at the university through conversations with ChatGPT.
This technology is called “Chip Chat”. Researchers only need to talk to ChatGPT to automatically write the chip design language Verilog that describes the chip function.
“By using the AI system associated with ChatGPT, the researchers hope to accelerate hardware design time to one month or less,” said Siddharth Garg, associate professor at the Institute of Electrical and Computer Engineering at New York University's Tandon School of Engineering.
Generally speaking, designing the most complex microchip may take up to half a year or even longer.
But these AI tools are not omnipotent.
Currently, these tools are mainly used to train young chip designers, write hardware languages and report errors, said David Pan, a professor of electrical and computer engineering at the University of Texas at Austin.
Current tools have other limitations. Engineers must carefully verify the output generated by AI, and there is currently no solution that can automate the entire chip design process from design to verification, implementing the design's transistors, and checking the design's electrical characteristics.
Synopsys' Krishnamoorthy estimates that the ability to autonomously create functional chips using generative AI is aboutIt will take 5 years.
References:
https://www.businessinsider.com/nvidia-uses-ai-to-produce-its-ai-chips-faster-2024-2
https://www.wsj.com/articles/designing-chips-is-getting-harder-these-engineers-say-chatbots-and-ai-can-help-092b4c49