In the 1980s, it was often said that "buying an IBM will never cost you your job." Even with the emergence of new technologies, IBM remained a safe choice. While it may not have been the most advanced option at the time, it was still undisputed. It had a mature ecosystem, and everything was under control.

Artificial Intelligence and EDA

When it comes to artificial intelligence, who or what is the safest bet? Who possesses the necessary data? Who has the right talent with the best knowledge and skills to create excellent solutions? Who has the funding? You might think of different companies for these questions, but the problem is that no company is willing to share more than necessary information. This means everyone can only access limited information.

For a long time, EDA companies have been tool suppliers in the semiconductor industry because they could achieve good economies of scale. But this wasn't always the case, and it may not be today. In the early days of the semiconductor industry, each company manufactured its own tools. There were no unified standards, and tools were often seen as a differentiating advantage. As standards emerged, each company realized that its tools were largely similar to those of other companies, making it less cost-effective to maintain its own. In many cases, they would transfer technology to EDA companies in exchange for years of free maintenance.

This situation may happen again. We've already seen systems and software companies transforming back into hardware developers, both at the semiconductor and systems level. Google has developed its Tensor processors for generations and has created several proprietary tools. It trained an AI model to understand Verilog code and its various design aspects, and has achieved a level where the AI-powered design system outperforms human designers. It has also developed test generation tools, some of which are already available for the RISC-V platform.

Nvidia's ChipNeMo

NVIDIA has developed a variety of tools, including ChipNeMo. According to NVIDIA's website, "Instead of deploying off-the-shelf commercial or open-source language learning models (LLMs), we employed the following domain-adaptive techniques: custom tokenizers, domain-adaptive continuous pre-training, supervised fine-tuning based on domain-specific instructions (SFT), and domain-adaptive retrieval models." The number of researchers involved in this work far exceeds that of any EDA company's engineering team I know of. This is just one example. Other work includes simulators, static timing analysis engines, place-and-route tools, and more.

So what about others? These tools are not sold to the public. Even if they were, it's unclear whether they would be equally suitable for designs outside their specialized training domain. Developing a tool that only needs to perform well on a small range of designs is much easier. EDA companies don't have this option.

While EDA companies have the most experience in developing and maintaining tools that meet mass-market needs, they lack all the necessary information, and may not even know what type of problems their internal tools are designed to address. Some companies may have limited knowledge in areas like high-speed interfaces, but they may have already developed tools that help them build these interfaces faster and better than their competitors. Publishing this information may not be in their best interest.

EDA and AI integration

This is why we're seeing more tentative progress in the convergence of EDA and AI. EDA companies are able to, and are developing agent solutions around their own tools. They provide APIs and MCPs, allowing customers to develop their own solutions, which may or may not integrate their internally developed tools, but they may not fully understand why certain requirements are being placed on them.

Researchers and startups are free to explore new ideas and possibilities, often without considering traditional systems, but this can make their solutions difficult to integrate into existing processes. They also lack sufficient financial resources to train models or support a large customer base, resulting in slow rollout - something no one can afford these days. This is why we're seeing tens of millions of dollars in venture capital flowing into EDA startups, hoping to achieve rapid growth. This isn't the first time we've seen massive investments pouring into areas that promise to disrupt industries. In the past, these investments have ultimately failed.

The current industry landscape is like the Wild West, with everyone struggling to figure out who is trustworthy, who can be partnered with, and how to meet industry needs. While we may be able to predict long-term industry trends (though we may not be able to predict which companies will ultimately succeed), short-term trends are difficult to predict.

Source: Compiled from semienginerring

Reference link: https://semiengineering.com/follow-the-ai-leader/

View more at EASELINK