On October 15–16, at the tail end of SEMICON West 2026, PDF Solutions held its 2026 CONNECT Conference at the Westin St. Francis in San Francisco.
PDF Solutions Vitals
Santa Clara–based PDF Solutions is a semiconductor data, analytics, and connectivity company, founded in 1991. Today, its business rests on three pillars, all feeding a single data layer:
- Exensio — the core analytics platform. It ingests 50+ data formats from FDC, test, assembly and packaging, harmonizes them into a semantic model (cloud or on-prem), and runs interactive analytics and ML on top. This is the recurring, scalable engine.
- Cimetrix + SecureWise — factory equipment connectivity and, since the March 2025 SecureWise acquisition, secure remote equipment connectivity for data acquisition and control. The connectivity layer is what makes PDFS a manufacturing data aggregator, not just an analytics vendor.
- DirectScan / Design-for-Inspection — on-chip electrical measurement hardware to catch defects during manufacturing. eBeam inspection to uncover buried systematic defects. This is the differentiated-data moat: PDFS generates proprietary characterization data at advanced nodes that optical inspection can’t see, then monetizes the analytics on it.
The company reports about 350 customers spanning fabs, fabless, OSATs and equipment OEMs, i.e., it sits in the middle and broker data across the chain. It is publicly traded on the NASDAQ stock market under the symbol PDFS.
The PDF Solutions CONNECT Conference
CONNECT was built around the recent announcement of Exensio Aurora and the requirements for applying AI across manufacturing, inspection, test and the supply chain, supported by customer deployment discussions. Its intended audience included manufacturing engineers, data scientists, equipment companies and semiconductor executives.
The agenda for Thursday morning, Oct 15, proposed two keynotes:
- Executive Keynote by John Kibarian, CEO, titled “Next Generation PDF Solutions Platform for the Global Semiconductor Industry.”
- Technical Keynote by Kimon Michaels, CTO, titled “PDF Solutions Platform – Architecture Overview and New Solutions Announcements.”
And an Exensio Aurora panel at noon.
Afternoon customer and solution talks: ST on advanced test solutions; Intel on reimagining manufacturing analytics for advanced packaging; Samsung on vector-scan eBeam inspection in advanced logic; Unimicron on advancing substrate manufacturing with FDC; PDF on its AI strategy and on transforming analytics with Exensio Aurora; POET on wafer-scale optics for the AI era; next-generation equipment integration and control; secureWISE updates. The evening dinner panel included the NYTimes, imec, Siemens EDA, Teradyne and PDF Solutions.
The Agenda for Friday, Oct 16, included an industry keynote by Dr. Jon Koomey, “Plenty of room at the top (and also the bottom): What’s next for computing,” with a fireside chat with Kibarian.
Then deep dives, clinics and demos on advanced test, scalable analytics, LLM-based solutions, equipment solutions, and Aurora development strategy and design.
Customer speakers and what they were billed to cover
- Samsung: Dr. Hyung Keun Yoo, Sr. Director Metrology and Inspection, on adopting PDF’s DirectScan technology to ramp leading-edge chips.
- Intel: Syed Baquar, Principal Engineer and Director, Data & Analytics, on how Intel is transforming its manufacturing data infrastructure and moving to AI-driven analytics with PDF.
- STMicroelectronics: Massimo Longaretti, Software Automation Manager, on how to automate test execution, moving towards “dark factories.”
- Unimicron: Dr. Justin Chang, Senior Deputy Director of Smart Manufacturing, on FDC for substrates.
- POET: Raju Kankipati, Chief Revenue Officer, speaking on managing operational challenges in a fast-growing company.
- Dinner panel: Teradyne CEO Greg Smith, imec CEO Patrick Vandenameele and Siemens EDA EVP Ankur Gupta, joined by Kibarian. The panel moderator originally announced was Bloomberg’s Emily Chang who in the end was travelling in Europe that week with Jensen Huang. Tripp Mickle, NYTimes tech reporter, very successfully moderated the panel in her place.
Main announcement: Exensio Aurora
- Unveiled September 9 as the new, highly scalable Exensio analytics architecture, with first public demos at CONNECT.
- PDF’s stated rationale is that 3D architectures, advanced packaging and heterogeneous integration blur front-end/back-end boundaries and cause a data explosion, and that in its experience under 5% of semiconductor manufacturing data gets analyzed with traditional tools.
- It was built because conventional BI tools hit in-memory processing limits and general-purpose cloud data platforms lack semiconductor relevance, at the urging of several customers. It is cloud-first but also supports on-premises deployment, for both online and offline applications.
- Said Akar, VP Enterprise Development, wrote a blog post on the four questions that led to Aurora, including how Exensio supports a range of AI approaches, LLMs among them.
- One trade outlet noted the caveat: the September announcement defined the architecture and direction rather than offering an independently benchmarked production result.
Executive and Technical Keynotes
CEO John Kibarian argued that AI could now transform semiconductor manufacturing itself. His executive keynote and co-founder Kimon Michaels’s technical presentation outlined how Exensio Aurora aims to connect manufacturing knowledge, scalable analytics and coordinated action across the supply chain.
Kibarian framed the opportunity around efficiency. Demand for AI hardware is growing, but equipment availability, energy and manufacturing capacity constrain production. New fabs require expertise and successful process ramps as well as capital. Customers must therefore extract more useful output from existing resources, even as they invest in additional capacity.
More products and more complex data
Specialized processors are changing manufacturing requirements. Foundation-model companies and data-center operators are developing bespoke chips, creating more products and potentially smaller production runs. Some customers, Kibarian observed, effectively never leave new product introduction: the extensive measurements needed during development remain important throughout production.
Assembly and test are also becoming more complicated. Fabless companies must manage information across multiple manufacturing partners and test stages. Earlier results can influence downstream screening, grading and routing. Advanced packaging, photonics and co-packaged optics add further complexity, making traceability and data integration increasingly important.
The challenge extends beyond collecting measurements. An experienced engineer understands relationships among products, tools, chambers and process histories that a general-purpose AI model may miss. During yield excursions, relevant warnings often become obvious only afterward. PDF Solutions’ strategy is to combine manufacturing data with semiconductor-specific context so engineers can identify meaningful signals sooner.
Reusable workflows connect AI with engineering practice
Kibarian described a practical role for large language models: constructing analytical workflows that engineers can inspect, execute and reuse. A request can generate blocks for querying data, performing calculations, producing plots and annotating results. Python provides the underlying language, while visual workflows expose the analytical sequence.
Once validated, a workflow can be stored and reused, reducing repeated model interactions and associated computing costs. It also preserves engineering knowledge. Manufacturers need consistent, auditable decisions, particularly when screening products with long service lives. Retaining the procedure makes it possible to investigate how a result was reached rather than relying solely on an AI-generated answer.
Michaels explained the architecture supporting this approach. Aurora shifts heavy analysis from desktop clients to distributed server-side computation. Its FUSE computing stack addresses both parallel processing and parallel data access, accommodating very wide development datasets and extensive high-volume manufacturing histories. Caching and reusable intermediate results further improve efficiency.
Michaels reported approximately 25-fold performance improvements in internal testing at comparable hardware spending. These are company-reported figures whose applicability depends on the workload. The broader objective is to make analyses practical that previously required excessive time or computing capacity.
Taking models into production
Exensio StudioAI provides the framework for model development, deployment, monitoring and governance. Tracking model versions and operating locations matters when an automated screening decision must be reconstructed months or years later.
Deployment depends on the application. Time-critical algorithms may run directly on testers or process tools. Factory-level analysis can combine information from multiple machines, while less urgent work can run centrally or in the cloud. Michaels emphasized an open architecture that supports customers’ own models, libraries, data stores and agents.
Exensio Aurora also expands the representation of manufacturing history beyond flat tables. Repeated measurements, retests, derived values and relationships among wafers, dies and assembly stages must remain connected. Integrating design and product-lifecycle information supplies additional context for analysis.
A detailed demonstration was provided of how LLMs can be used to “dialog with the data” and create successive analytical workflows that can be saved and reused later for similar purposes ensuring repeatability, predictability and efficiency while encoding the specific engineering knowledge of the users.
Kibarian linked these capabilities with secureWISE, which supports controlled, encrypted collaboration between fabs and equipment suppliers. PDF Solutions is extending that connectivity toward fabless companies and outsourced assembly and test providers. Sapience Manufacturing Hub adds connections between manufacturing information and enterprise processes, helping analytical insights inform operational and business decisions.
A successful event with a clear direction
CONNECT 2026 succeeded in giving PDF Solutions’ strategy a concrete technical foundation. Kibarian established the manufacturing problem; Michaels explained the architecture intended to address it. Together, the keynotes presented a coherent path from better data to reusable analysis and coordinated action.
The event’s lasting message was practical: AI’s value in semiconductor manufacturing will be measured by faster diagnosis, better equipment utilization and production decisions that engineers can understand and trust.
Also Read:
Who Gets to Move the Recipe? How Autonomous Agents Earn Decision Authority in the Fab
Share this post via:
Comments
There are no comments yet.
You must register or log in to view/post comments.