AI Takes a Strong Supporting Role in Modern Chip Design
With AI evolving, chipmakers are taking advantage of deep learning methods to design chips faster and more efficiently than humans.
Integrated circuit (IC) design is a complex endeavor, ever pushing the boundaries of density and performance. Now, artificial intelligence (AI) is increasingly playing a role in IC design.
While AI has recently gained more attention for its surprising abilities with natural language processors like ChatGPT, this tool is also being used in various stages of IC design, including design optimization, layout, simulation, and verification. AI algorithms can also help explore design space more efficiently, discovering optimal design configurations faster than traditional methods.
AI can be a strong asset to IC designers as designs become more complex than ever. Image (modified) courtesy of Synopsys
This article highlights how AI has been used as a tool for IC design and how this technology may affect the professional value of IC designers.
Chipmakers Leverage AI for Strict Design Constraints
With billions of transistors confined to a small chip area, IC designers are tasked with optimizing design despite strict constraints.
For one, the die area must be minimal to fit the small form factors of devices today and keep manufacturing costs down. The layout's power consumption is also of concern; power likewise affects the cost of deployment along with the chip's environmental impact. Certain dense regions and configurations are prone to overheating, requiring cooling mechanisms or clever layouts. Given these and many other factors, IC designers spend roughly eight to nine months to generate a chip layout that addresses each stringent requirement.
To speed and optimize the IC design process, many companies—including some of the biggest in the tech industry—are now investing in AI tools to do some of the heavy lifting.
Google AI Designs Chips in a Few Hours
Google's deep learning reinforcement learning (RL) method can generate effective layouts in far less time than humans, and the results are comparable in quality, according to Google. In March 2022, Google Research introduced PRIME, a deep learning approach that uses existing data like power and latency to create accelerator designs that are faster and smaller than chips designed with traditional methods.
Google's PRIME implements logged accelerator data to train a conservative model for design accelerators. Image courtesy of Google
Google researchers used 10,000 chip floorplans to train their model. The AI-generated chip took less than six hours to design. The Alphabet company says this method has already been utilized to implement Google's tensor processing units (TPUs), a part of Google's cloud-based machine learning applications.
EDA Companies Double Down on AI Design Systems
It is not just Google turning to machine-learning models, either. Electronic design automation (EDA) companies like Synopsys and Cadence also use AI techniques in their newest tools. Recently, Synopsys registered 100 commercial tape-outs with its Synopsys DSO.ai autonomous chip design system. Recent customers of this system include STMicroelectronics and SK Hynix.
DSO.ai case study with no prior learning (left) vs. DSO.ai case study with prior learning. Image courtesy of Synopsys
ST and Synopsys used DSO.ai for the first time at the beginning of February 2022 on Microsoft's cloud to design a working chip. Using Synopsys' DSO.ai design system, combined with Synopsys Fusion Compiler and IC Compiler on Microsoft Azure, the tool increased power, performance, and area (PPA) metrics by more than 3x with up to 25% lower total power and a significantly smaller size.
Beyond chip design, AI also finds applications in chip testing and verification—both areas chip manufacturers spend much of their time. To address this stage of design, Siemens unveiled Questa Verification IQ, a software platform to help IC design engineers speed up the verification process.
NVIDIA's GPU-powered AI for GPU Design
NVIDIA devised another deep-learning approach for chip design. The company made an RL model called PrefixRL, proving AI can learn circuit design from scratch and make smaller and faster circuits using the latest EDA tools. NVIDIA's architecture consists of 13,000 circuits designed using AI technologies.
Depiction of the PrefixRL flow. Image courtesy of NVIDIA
For automated cell migration, the company developed NVCell, which can migrate 92% of the cell library with no errors. Humans can work on the remaining 8% of the cells that didn't migrate automatically. NVIDIA's chief scientist Bill Dally said:
"So this is like an Atari video game, but it's a video game for fixing design rule errors in a standard cell. By going through and fixing these design rule errors with reinforcement learning, we're able to basically complete the design of our standard cells."
Impact of AI on IC Design Job Market
While AI can automate certain tasks traditionally performed by IC designers, such as layout design and optimization, it also reduces the need for manual aspects of the design process. While this increases overall efficiency, it may also lead to eventual job displacement in certain areas of IC design.
On the other hand, AI can also help IC designers perform their jobs more efficiently and effectively. For example, AI can analyze large amounts of data and provide insights that suggest design alternatives an engineer may not have considered before. This trend can enhance the value of IC designers in the industry by allowing them to focus on more complex and creative aspects of design and ultimately produce better products.
It's unlikely that AI will fully replace the need for skilled IC designers. The demand for such engineers may even increase as AI becomes more prevalent in the industry, as there will be a need for individuals who can accurately verify and utilize AI tools and algorithms in the design process.