This article is compiled by Semiconductor Industry Review (ID: ICVIEWS) from semiengineering.While everyone seems to agree that artificial intelligence will disrupt semiconductor design and EDA tools, no one has proposed what the disrupted process will actually look like.
In recent years, artificial intelligence (AI) and machine learning (ML) have made breakthroughs in many application areas. In the traditional semiconductor field, researchers have also explored new methods for chip design based on machine learning. These new algorithms often first reflect in chip design tools, commonly referred to in the industry as ML for hardware design, or ML for EDA (strictly speaking, the former has a broader scope). Sometimes it is also referred to as intelligent EDA algorithms or intelligent IC design methods.
Anand Thiruvengadam, Senior Director and AI Product Management Lead at Synopsys, stated: “AI has the potential to change how customers design chips. The entire EDA process could be disrupted by AI.”
Chips are ubiquitous in our lives. The design and implementation of chips involve a complex process.
Driven by Moore’s Law, the increase in complexity means compromises must be made. Most of these compromises manifest in terms of creativity. Incrementally expanding on known foundations is faster, cheaper, and safer than starting from scratch each time.The introduction of IP further solidifies this concept and reinforces the need to lock legacy software into hardware architectures. This is entirely a design bias.
The rise of parallel processing is not because people think it is good, but because single-processor architectures have reached their limits. It took another decade or more for it to be widely adopted, and the introduction of early machine learning technologies became a significant event for its real development. If there had been an AI-driven chip design tool at that time, trained on the collective wisdom of the entire industry, could it have achieved this leap on its own? The author is seriously skeptical.
While it should understand parallel processing (which was more common in the 1980s than in the 2010s), and it should know how to write code to implement parallel processing, the bias towards single-processor design would leave it at a loss. It would learn how to use multiple single processors, which typically act as agents contributing small functionalities to the overall function (such as audio processors, USB controllers, graphics controllers, etc.), rather than using a central heterogeneous processor that can handle everything more efficiently.
Did NVIDIA initially intend to create an AI processor? No. They gradually met customer needs. It wasn’t until computer vision had enough datasets for training, along with suitable hardware, that deep learning was first proven to outperform rule-based systems. This led to many advancements we see today.
The idea of training semiconductor AI systems with all existing data will never be realized. Each company focuses on specific types of chips and specific types of problems, such as mobile, automotive, data centers, etc. Their biases are even more entrenched. Few hardware companies are also software developers, even if they may need some software that relies on hardware.AI will not be asked to invent something entirely new, but rather to make incremental improvements to hardware or software.
What would disruption in the semiconductor industry look like? It will not happen all at once, nor will it happen all at once.The author expects that the disruption of EDA will first appear in areas like high-level synthesis (HLS), where tools can be trained on a large number of architectures. This training already has enough data, and companies can enhance it. This will enable it to receive specifications similar to English and generate code, which can then go through traditional EDA processes.
Although this has been the goal since the inception of HLS, it has proven too difficult or only achievable in limited applications.SystemC has not yet been truly accepted as an input language, and even so, tools require a highly constrained language structure. But if AI can help achieve this goal, when custom designs are available for a larger community, the user base could expand by 10 times or even 100 times. This is enough to change the entire EDA process, which is why this technology received such massive investment when it first emerged.
Virtual prototypes also need to be able to operate according to these specifications to identify errors or omissions, and significant improvements in sequential equivalence checking are needed to ensure confidence in AI transformations. Over time, new core processes will introduce more agent assistants to handle issues such as power consumption and cost.
Disruption occurs when certain things change, not when certain things are optimized.
*Disclaimer: This article is the original work of the author. The content reflects their personal views, and our reposting is for sharing and discussion purposes only, not representing our endorsement or agreement. If there are any objections, please contact us.
