Smith: What trends are you seeing in the semiconductor industry. Are there any that surprise you? 

Kibarian: We see several trends that have been going on for quite a while.

As much as we hear Moore’s Law is dead, there’s still a strong drive to get to advanced nodes. The benefits are harder to achieve and require more than geometry scaling, but demand for these advanced nodes continues to grow. Another emerging trend is the need for insatiable compute power in data centers to support the explosion in AI applications. In recent history, the mobile phone market has been the key driver of the push to new advanced nodes, but that is changing as the performance needs of data centers and AI applications are now driving the shift.

Next, as companies are still learning from the disruptions in the supply chain due to the pandemic, there’s a tremendous amount of movement to make the supply chain more resilient by expanding sourcing options for critical products or test applications. This is happening in conjunction with significant investment in high-performance compute from many countries that want to bring silicon to their shores.

The next trend is that electronics companies are looking to limit investing solely in China or the U.S. Their China Plus One or U.S. Plus One strategies results in adding significant additional infrastructure and overhead. If it’s not done right, it will cost the industry more money. It will be hard to sustain the cost benefits and economies of scale of the current single source model just by brute force and adding human capital. A new approach is required to manage cost effectively smaller and globally distributed manufacturing facilities.

The final trend is the general electrification of the economy. Cars are moving from internal combustion engines to electric. That means more and more of our energy needs are met with electricity, putting a premium on solar and batteries. Batteries require power conversion.

Silicon such as high bandwidth semiconductors on silicon carbide and gallium nitride have a tremendous amount of capacity. What is interesting is how fast and aggressive China is in that part of the market; they could be a major producer of the technologies needed to support electrification. With our exposure to the China market as well as the European and U.S. markets, Chinese manufacturers have come up quickly, and we may see a world with more viable suppliers than originally anticipated.

Smith: You mentioned data centers and AI. AI is everywhere and revolutionizing the semiconductor industry. EDA companies are talking about incorporating AI. What are you observing? 

Kibarian: AI is used for chips that are manufactured for use in data centers. For example, our customers use PDF analytics or the Exensio platform via the cloud to analyze large amount of manufacturing data and product or test engineering data. Without this type of automated solution, only a small proportion of these data sets would actually be utilized.

Companies staff their product design and test engineering using a budget based on a percentage of revenue. If a company has billions of dollars of revenue, it will put so much more into product and test engineering. But how productive can these people be? Without AI, they can only use some simple reports and graphics to analyze the subset of data they are looking at. AI solutions such as PDF’s Guided Analytics capability apply sophisticated machine learning tools to analyze entire large data sets. AI is enabling engineers to be more productive by allowing them to work with large data sets that ultimately deliver better results in the products.

The amount of compute keeps going up at a rate that outpaced the rate of geometric scaling. More compute power makes it cost effective to go through large data sets and identify what is relevant.

Additionally, AI is helping semiconductor companies build products. A conventional compute system is chips assembled on boards. AI is making system-in-package take off.

The production flow is more complex, as fabless companies are becoming system companies. Conversely, system companies are becoming fabless companies and manufacturers. In the past, they ordered parts from their foundry of choice. Essentially, the foundry was the system manufacturer, supplying package and test yields of 99%.

Now companies are building systems in more complex packages potentially with foundry partners, but this requires getting known good die. High bandwidth memory or other components from other suppliers means the company must make sure these products are available at the right time. In essence, they are becoming manufacturers and changing the way customers manage the problem of product test. They’re adding more test insertion points and using machine learning and AI to be more productive.

Smith: Let’s talk about digital twins or creating virtual models of everything from chips to the whole system. How do you see the impact or effectiveness of digital twins in manufacturing? 

Kibarian: From a manufacturing perspective, digital twins had been models for chamber behavior on a processing tool like an etch tool or TCAD simulation of devices and structures.

The problem is that purely physics-based digital twins don’t exist, and we must utilize empirical data. The joke was that the modeling for tomorrow’s systems was based on yesterday’s technology. Trying to have the physics catch up with the materials, device structures and behaviors is why it’s so expensive to develop new technology.

Principles-based models will never catch up with production. We can model 90-nanometer technology, but it doesn’t work for one or two nanometer wafers. AI and machine learning – and ways of building models using more sophisticated algorithms – can help close that chasm, and that’s starting to happen at the R&D level.

In production, no one has yet achieved a good merger of the physics-based and AI-modeling worlds to create a virtual model. Virtual modeling is a big opportunity.

The rate of change and improvement in algorithms in large language models moves fast because machine learning can scrape the Internet for data to build huge training sets. In the semiconductor world, however, data sources are typically siloed within organizations and often not shared with vendors. This limits the rate at which the industry can take full advantage of existing data and create tangible economic benefit.

By and large, there is a lot of wasted capacity in semiconductor manufacturing. The operational effectiveness of factory equipment is up to 90-95%. The reality is that most factories today process product wafers 40-60% of the time – maybe 70-75% of the time on a test floor. It is critical for the industry to start leveraging new types of AI models to increase the productivity of its manufacturing capacity.

The industry needs to look at how companies can share data to take advantage of more sophisticated AI and create a new kind of operational digital twins. If the industry doesn’t make a change; it will only be the largest facilities with the largest datasets able to take advantage, leaving one or two winners, with the others not being competitive.

Smith: Is it possible for the industry to come up with a standard or some way of sharing information to build better models without giving away the underlying proprietary data? 

Kibarian: We can look at computer science with technology like homomorphic encryption. The relationships between parameters remain, but the underlying numbers or raw data is not visible after encryption. Pharma and the medical industry have ways to add noise to the data while preserving the information, as required by the Health Insurance Portability and Accountability Act (HIPAA).

Our industry has a knee jerk reaction when it comes to looking at how to take full advantage of data and prefers to solve it as if information and data is more proprietary than medical data or financial data. And I don’t think that’s true.

Bob Smith: Is the open-source movement destined to bring change to the industry? 

Kibarian: PDF is a big believer in open source when it comes to OS-level virtualization and Kubernetes versus proprietary alternatives. We also use open-source database technology like Cassandra but are skeptical of the value of open-source solutions for end-market verticals. Having an underlying open and available IT layer has tremendous value, because it means a more rapid rate of innovation and greater ability to adjust security vulnerabilities and patches versus proprietary systems.

Smith: PDF sits right between manufacturing and design. On the EDA side, more collaboration is going on between designers and manufacturing. How would you bring these two domains closer together? 

Kibarian: That’s a good question. My first instinct is to look at the largest design organizations and manufacturers. They often invest heavily to figure out how to get jobs done right. This results in the concentration of the industry on a smaller number of players and leads to less innovation. However, in the world of chiplets and advanced packaging, there are more opportunities to become a chiplet supplier, because the whole system doesn’t need to be built by a single company. A supplier of chiplets could sell it into many systems

From a system view, connecting the pieces together through software, data sharing and analytics could drive more productivity gains that will offset some of the natural headwinds. This needs to be addressed in a way that changes the paradigm with software and systems used to bring manufacturing and design closer together.

About John Kibarian

John K. Kibarian is President, Chief Executive Officer and Co-Founder of PDF Solutions. He has served as President since 1991 and CEO since 2000. Dr. Kibarian received a Bachelor of Science degree in Electrical Engineering, a Master of Science and PhD degrees in Engineering Computer Science from Carnegie Mellon University.