IN Brief:
- Exensio Aurora is designed around semiconductor manufacturing datasets operating at petabyte scale.
- PDF Solutions intends to combine parallel processing, machine learning, and AI with semiconductor-specific analytical workflows.
- The architecture links online manufacturing functions with offline analytics, although public performance demonstrations are scheduled for October.
PDF Solutions has introduced Exensio Aurora, a new architecture for its semiconductor manufacturing analytics platform designed around petabyte-scale datasets, parallel processing, machine learning, and agentic AI. The company plans to demonstrate Aurora at its CONNECT event in San Francisco on 15–16 October, so the current announcement defines the architecture and development direction rather than providing an independently benchmarked production result.
Semiconductor manufacturing creates a difficult analytical problem because the data is not merely large. Fabrication, assembly, and test generate information from process tools, metrology, inspection, parametric test, electrical test, equipment sensors, manufacturing execution systems, yield analysis, and increasingly complex packaging flows. Engineers may need to relate a result from one die to its wafer, lot, process history, equipment state, assembly route, and downstream test behaviour.
PDF Solutions says Aurora has been designed to handle that variety at petabyte scale during analysis as well as storage and ingestion. Its previous platform evolution concentrated heavily on accommodating larger datasets; the new architecture is intended to bring parallel processing, automated analysis, machine learning, and AI into the same environment rather than treating them as separate additions around a conventional database.
The distinction matters because storing more manufacturing history does not automatically shorten an investigation. Yield engineers routinely work backwards from a failing population through test bins, wafer maps, process conditions, tool histories, and other contextual data. If large datasets still have to be manually reduced or exported before meaningful analysis begins, the storage platform can become another staging area rather than an engineering tool.
PDF Solutions also argues that general business-intelligence systems make assumptions about data shape that are poorly suited to semiconductor production. Manufacturing datasets can be extremely wide, sparse, hierarchical, and inconsistent between process steps, while advanced packaging creates additional relationships between several dies and assembly operations. The data model therefore has to preserve semiconductor context rather than flatten every measurement into a generic enterprise schema.
Aurora is intended to span online and offline manufacturing use cases. Online systems sit close to equipment and process control, where data has to be collected and acted upon quickly. Offline analytics works across longer histories to identify excursions, correlations, yield losses, or systematic patterns. PDF Solutions’ argument is that separating those environments weakens the feedback path between what engineers discover and what the factory subsequently measures or controls.
That feedback becomes more important as machine-learning models move towards production equipment. A model trained on historical fab data can potentially identify conditions associated with an excursion, but deploying it near the process requires controlled interfaces, version management, monitoring, and evidence that the output remains valid as products and recipes change. Semiconductor manufacturing has little tolerance for an opaque model quietly drifting while high-value wafers continue through the line.
Agentic AI raises the bar further. PDF Solutions explicitly acknowledges the non-deterministic behaviour of large language models, which is a poor match for process decisions that need repeatability and traceability. The useful role for an AI agent is therefore likely to depend on how tightly it is constrained by semiconductor-specific data, permissions, validated analytical functions, and human engineering review.
That is a less glamorous problem than adding a conversational interface to a dashboard, but it is considerably more important. A system that suggests a root cause has to expose the evidence behind that suggestion, while a system capable of taking action has to distinguish between exploratory analysis and changes that can influence production equipment or process control.
Aurora’s public demonstrations in October should provide more evidence about how those controls, workflows, and performance claims operate in practice. Until then, the architecture is best treated as PDF Solutions’ next platform generation rather than proof that agentic AI has solved semiconductor yield analysis. Petabyte-scale fab data is already a reality; turning it into repeatable engineering decisions without losing provenance is the harder problem Aurora now has to demonstrate it can address.
