The semiconductor manufacturing industry faces an unprecedented data challenge. For the newest devices, test programs can contain over a million test items, generating gigabytes of data per chip across probe, assembly, and test operations. The largest deployments have reached the multi-petabyte range, creating a fundamental problem: traditional business intelligence tools simply cannot handle semiconductor-scale data with millions of columns and rows.
Public comments from three semiconductor executives sum up the challenge. “As a result of the increased complexity of advanced packaging, the amount of manufacturing and test data that semiconductor companies need to analyze has increased sixfold since 2022,” recently commented Mike Campbell, Qualcomm’s chief supply chain Officer.
At the same event, Aziz Safa, corporate VP and GM of Intel Foundry Automation, had this to say: “We have 600 petabytes of data across Intel. The challenge that we have is to be able to run algorithms in the areas where we need that data to solve problems.”
And John Kibarian, CEO of PDF Solutions, mirrored those remarks. In many cases, he said, no more than 5% of the collected semiconductor manufacturing data is used in analytics. Yet more than ever, access to timely analytics is critical to quickly ramp the yield of new advanced process nodes or ensure the quality of complex packages. In this context, it’s critical to find new innovative ways to scale the ability to analyze semiconductor data.
One comprehensive strategy includes a plan to enhance the capability of a data platform, already widely used across the semiconductor industry, to address this challenge by combining scalable analytics infrastructure with advanced AI capabilities, including large language models (LLMs) and autonomous agents.
This approach represents a fundamental rethinking of how semiconductor manufacturers can extract actionable insights from massive, complex datasets.
The scalability problem
Traditional business intelligence (BI) tools face critical limitations in semiconductor manufacturing environments. They rely on local memory, which severely restricts analysis and machine learning capabilities. They also lack computational and organizational scalability often related to the specific characteristics of semiconductor data that may have hundreds of thousands or even millions of parameters to analyze.
Think of a table with a million columns and hundreds of thousands of rows. Visualizing this type of dataset in a traditional data analytics or BI tool has reached its limit, and this approach will not address the future needs of an industry where data size and complexity keep increasing.
Typically, engineers develop bespoke scripts based on summary statistics disconnected from the original data sources, and these scripts are typically served without the infrastructure for robust sharing across the organization.
One answer is a new parallel and distributed data architecture with dynamic partitioning. Rather than bringing raw data to the client for analysis, the system keeps data in the server layer and delivers only the visualizations needed by users. This thin-client approach enables the system to scale dynamically based on current needs by caching in the data layer for faster access and pre-configured analytics running continuously across all available data.
The results are striking. Benchmark testing shows approximately 25-fold performance improvements on typical large test programs with the ability to work with one million test items and beyond, a scale of analysis previously impossible.
The system achieves this through parallelizable performance across both rows (individual die) and columns (test parameters), combining static compute nodes with burst cloud computing for cost-effective scaling to extremely large datasets.
Deploying AI models at scale across enterprise
Deploying AI in semiconductor manufacturing requires more than just training models; it demands a complete operational infrastructure. The infrastructure’s architectural strategy addresses three major operational challenges: deployment bottlenecks caused by manual handoffs and brittle integrations; data friction from building custom pipelines instead of leveraging existing systems; and governance risks from poor lineage between production models and training parameters.
One tool gaining market traction used by data scientists from code to production for semiconductor data is focused on deploying models at the edge. Add-on capabilities include the ability for engineers to add their own models.
An enterprise-grade model registry will enable model lifecycle governance, tracking, and sharing, with full data traceability ensuring that any model’s training inputs are always known.
Breaking down data silos
One of the most significant challenges in semiconductor manufacturing is the fragmentation of critical data across isolated systems. Yield data sits in one place, design diagnosis information in another, and equipment telemetry in yet another. This fragmentation blocks the correlation of volume yield data with physical layout features and prevents engineers from connecting specific process excursions with final yield outcomes.
One solution is extensive data integration efforts via a platform extending beyond traditional manufacturing analytics supported by a semiconductor-specific end-to-end data model.
Central to this effort is the development of a semiconductor-specific semantic data layer that maps the complex relationships between yield, design, process, and tool data. This allows alignment and linking data across domains and sources in the data platform. It also allows LLMs to interpret disparate data types as a unified whole rather than struggling with disconnected information sources.
Workflows as the foundation
A key architectural decision in the platform is to treat workflows as the internal language of the system. Every analytic operation—whether rules, machine learning pipelines, or batch analytics—is expressed as a workflow.
This provides several critical benefits. Workflows serve as the long-term memory of the system, capturing not just results, but the complete methodology used to achieve them. They can be created from learn mode, through LLMs, manually, or programmatically, and can be embedded within larger workflows for maximum reusability. Engineers may never need to directly interact with a workflow, but the capability is there when needed.
Critically, workflows act as semiconductor-specific content and context, encoding best practices as reusable playbooks. They provide transparency into how results are achieved and serve as guardrails for AI reasoning, helping prevent the hallucinations that can occur when LLMs operate without domain constraints.
The agentic LLM platform
The goal is to enable engineers to interact with manufacturing data at a higher level of abstraction. Rather than requiring deep technical knowledge of query languages and data structures, the result is a system where engineers can ask natural language questions and receive actionable insights.
Achieving this vision requires a “Semantic, Agentic, and Secure” infrastructure. The semantic layer is built on domain expertise, creating semiconductor-native knowledge graphs that encode the fundamental data hierarchy of manufacturing. This anchors LLM reasoning in the structural reality of manufacturing data, eliminating ambiguity and providing the ground truth context needed to prevent hallucinations.
For example, the system understands that CV refers to Characterization Vehicle, that yield represents the results of die binning, and that the data hierarchy flows from lot to wafer to die to package. It knows that common analytical tasks include yield trending, bin Pareto analysis, and univariate screening. This enables engineers to ask questions like “Show me the yield trend over the last week” or “What is the root cause of low yield in lot XX?” and receive meaningful, accurate responses.
The platform integrates a model context protocol for a truly agentic system. Rather than just summarizing text or answering questions, the system can autonomously plan and execute complete workflows from raw data ingestion through complex plot generation.
To ensure reliability and transparency, any agentic tasks are executed using scalable analytics workflows. They can be viewed, saved, and modified by engineers at any time to ensure total transparency of LLMs actions.
To ensure the sensitivity of semiconductor manufacturing data, a fully air-gapped, on-premises LLM infrastructure option, designed for intellectual property sovereignty, can be added. This ensures that sensitive yield data and proprietary models never leave secure firewalls, eliminating reliance on third-party cloud providers.
The path forward
A platform like this requires thorough research and development on technology selection, validation and tuning, engaging a large group of architects, developers, quality assurance specialists, designers, and product managers.
This type of platform addresses the critical industry challenge: de-risking AI adoption by securely scaling execution and maximizing return on investment from legacy data, while simultaneously future-proofing infrastructure for the rapidly emerging age of LLMs and autonomous agents.
By combining massive-scale data processing, an operational enterprise, intelligent data integration, and agentic LLM capabilities, all grounded in deep semiconductor domain expertise, the industry can be transformed. The platform can identify how value is extracted from the exponentially growing volumes of manufacturing data.
The approach suggests a future where engineers spend less time wrestling with data infrastructure and more time solving the complex yield and quality challenges that define success in semiconductor manufacturing.
Peter L. Kostka is a Vancouver-based technology entrepreneur with a track record of scaling complex deep-tech concepts into successful commercial outcomes. Currently, he serves as the director of product management for AI at PDF Solutions, where he spearheads the AI technology roadmap and leads rapid prototyping for semiconductor and battery manufacturing sectors.
Editor’s Note
Presentations by Qualcomm’s Mike Campbell (“AI-Driven Innovation in the Semiconductor Industry”) and Intel’s Aziz Safa (“Enabling AI/ML strategy using the PDF Suite”) were given at the 2025 PDF Solutions Users Conference.
John Kibarian’s “Revolutionizing Semiconductor Collaboration: The Emergence of AI-Driven Industry Platforms” keynote was presented at SEMICON West 2025.
Related Content
- Coming Soon: Low-Cost Mini Fabs
- 5 ways manufacturers benefit from AI in chip design
- Tapping AI for Leaner, Greener Semiconductor Fab Operations
- Addressing the Biggest Bottleneck in the AI Semiconductor Ecosystem
- Unlocking compound semiconductor manufacturing’s potential requires yield management
The post A real-world approach for AI-driven semiconductor manufacturing appeared first on EDN.
