Posted on semi.org Click here to view original article
By: Bob Smith
John Kibarian, CEO and co-founder of PDF Solutions and a member of the ESD Alliance (ESDA) Governing Council, will deliver a keynote during the CEO Summit at SEMICON West in October titled, “Revolutionizing Semiconductor Collaboration: The Emergence of AI-Driven Industry Platforms.” He recently shared with me a summary of what his talk will cover and his perspective on why collaboration is the key to growing the semiconductor industry to $1 trillion and how we will get there.
Smith: What is the major industry evolution or dynamic that’s making collaboration essential today?
Kibarian: The semiconductor industry has dramatically evolved from a simple, linear process to a complex, collaborative ecosystem. Previously, everything happened at the wafer fab. Testing occurred at wafer sort, package yields were high, final testing was straightforward, and products were shipped. Collaboration was mainly between foundries and fabless customers, intensive during early qualification and test chip stages, then evolved to routine yield monitoring once production stabilized.
Today’s advanced packaging puts multiple chiplets into single packages, creating an explosion of test insertion points. This has driven up both test complexity and costs significantly. Front-end fabs now house the most complex machines, while test and assembly facilities, once relatively simple, now feature sophisticated system-level testers with integrated robotics. Assembly tools have become highly complex, with die-attach processes requiring increasingly tight tolerances.
Success now requires collaboration across the entire chain, from system companies to equipment vendors, both for new product launches and ongoing production maintenance.
Companies are adopting AI and machine learning (ML) to manage these complex production flows, whether for testing or equipment control. This demands even broader collaboration since AI requires combining data from multiple sources across foundries, fabless companies, OSATs, equipment vendors, and more, data that no single entity controls. What was once a straightforward handoff between two parties has become an intricate web of interdependent relationships requiring continuous coordination.
Smith: Chiplets and chiplet-based architecture is in the news and seems to be a key solution or practical solution in response to the slowing down of Moore’s Law. This demands incredible levels of collaboration and coordination across the whole value chain. Is this doable at scale where it starts to move into the mainstream?
Kibarian: The semiconductor industry will need unprecedented collaboration to make chiplet manufacturing work at scale. But this can be done! Consider EUV lithography: Initially expected during the 65nm generation, it took years longer than anticipated despite being an incredibly complex technology. Extraordinary engineering was needed but it also required extensive collaboration between ASML, suppliers, customers, and the broader fabless community.
If the industry achieved this level of coordination for EUV, it can do the same for chiplets. However, chiplet manufacturing will require even greater collaboration as more companies will build systems using chiplets from multiple suppliers.
Today’s chiplet-based systems typically source all components from one manufacturer, making standards like UCIe less critical since companies control their entire supply chain. This will change as companies increasingly use third-party components for cost-effectiveness.
More and more, we will see systems using components from multiple players to get to market more cost-effectively. Consequently, future production flows will be significantly more complex, requiring coordination of substrates and base dies, third-party dies and interposers, OSAT and specialized testers with specific configurations.
This orchestration must work not just for initial bring-up but for the ongoing production as well, and when reconfiguring chiplet combinations for different products, all requiring rapid, automated responses.
All of that must be automated for quick reaction. Considering the complexity of the manufacturing flow, people will want to apply AI/ML to anticipate what is going on in each individual product built.
Manual oversight of every chip and package during manufacturing isn’t feasible at scale. Automated AI agents must handle this monitoring and quality control. Expanding this automation will require close collaboration between the manufacturing entity and engineering teams at the product companies.
This will also require a different level of alignment and orchestration across all the software packages managing this complex multi-company process. The financial enterprise resource planning (ERP) systems know where material is going, what the demand is, and what the forecasts are. While separately, the manufacturing execution systems needs to know which tools are going to be available when.
Most often, these manufacturing systems operate in factories the product company doesn’t own. The product company’s PLM systems control the bill of materials and test flows, but these tests will be conducted at the OSAT requiring complex coordination between the software systems of multiple companies controlling different process domains. This orchestration spans organizational boundaries and must be able to take data from upstream test results and make decisions on what tests to run downstream. This is required to get the right chiplets put together into a package in an efficient manner within a short cycle time and not require a Formula 1 pit team to keep everything running.
Smith: The volume of data is staggering, especially now with design data. What will it take to enable this vision, at scale, where everything’s connected?
Kibarian: It’s a marriage of the human establishing the bounding box within which the systems operate that employ agents to do a lot of the work on a day-to-day or hour-by-hour basis.
A good example is how manufacturing execution systems (MES) connect to ERP systems to share data. When a company sets up an orchestration, it creates rules that govern how information flows between systems. These rules tell the ERP system: “To calculate costs for each process step, here’s the recipe information you should use.”
Once these rules are in place, they work like guidelines that control daily operations. An AI agent automatically creates insights based on actual data collected from the MES and moves data between systems according to these rules. The ERP AI agent will use this data to spot when costs are rising and send alerts, to notice when production yields drop, to calculate what lower yields mean for costs, and will take action to fix problems.
This same process happens between equipment suppliers and manufacturing facilities. They share data automatically based on pre-set rules, and AI helps identify issues and take corrective action. Fabs determine who can access which machines and when, what types of data can be transmitted and through which channels, and how frequently these transmissions occur. When new software or AI models are introduced to run equipment, the systems specify what virus scanning and security checks must be completed before installation.
Human operators primarily configure these control systems by determining the most effective collaboration protocols. However, the day-to-day execution is handled by automated agents due to the enormous scale involved, both in terms of data volume and the sheer number of transactions that occur continuously throughout operations.
A human will not go through and review that data. I’ll give two examples of this. One outside of our industry and one in our industry.
At our 2019 user conference, board member Marco Iansiti, a Harvard Business School professor, shared insights from his book on AI in business. He compared traditional banks with Ant Bank, Alibaba’s banking arm, which was experiencing explosive growth before Chinese government intervention.
Ant’s AI wasn’t particularly sophisticated, but its process was revolutionary. While traditional banks require customers to fill out loan applications that then go to human loan officers for review, Ant’s system would automatically scrape the internet and social media to verify applicant information. Within seconds, an algorithm would approve or deny the loan.
The crucial difference is that Ant could scale exponentially because its only constraint was computing power. Traditional banks need to hire more loan officers to double their business, a human bottleneck that limits growth.
I invited him to speak because I believed in this principle six years ago, and I’m even more convinced now.
For the semiconductor industry, to build a trillion-dollar industry with complex, integrated systems, we need to minimize human intervention in data intensive processes. Despite the trust issues between stakeholders in our sector, collaboration remains essential. The solution requires establishing systematic principles that allow AI agents to operate autonomously. This is a way forward to achieving exponential growth.
The Ant Bank example perfectly illustrates what our industry needs. At PDF, we believe this approach is crucial for industry advancement. Consider this: We manage petabytes of data, yet humans only examine 5-10% of it. This shows AI’s potential to handle the vast majority of operations without human oversight.
The reality is that our customers build millions of chips a week, billions a year. They cannot look at every dataset. Algorithms can, AI can. We launched a product called Guided Analytics last year. An engineer spoke about it during our user group last year. Her company has a couple of thousand products. Her group could not keep track of them every day, but Guided Analytics could. When her group came in the morning, the daily report noted 90% of the chips were fine or alerts pointed to where issues are. It’s a simple AI bot crawling over data and identifying where the root cause seems to be.
Our industry will require more agents to scale. Those agents will span the industry, and yet we as humans need to set up the governing principles under which they can operate. That’s how we’re going to deal with the massive amounts of design and manufacturing data to get the velocity the industry will need, and to benefit from the AI that we create for our businesses.