As semiconductor devices grow in complexity and chiplet-based architectures become mainstream, traditional static test plans are increasingly inadequate. To enhance test efficiency and accuracy, dynamic, data-driven strategies — including adaptive test limits, predictive analytics, and real-time decision-making — are becoming essential. A growing trend is the use of data feedforward and data feedback across test insertions (e.g., parametric test, wafer sort, final test, and system level test), powered by ML (machine learning), to enable optimization techniques like predictive binning, shift-left testing, and multivariate outlier screening.
However, implementing data-driven test optimization faces three major challenges, especially within the fabless-OSAT (Outsourced Semiconductor Assembly and Test) manufacturing model. First, model training requires aggregating data from multiple test insertions, which are often geographically distributed across different factories. Second, real-time model inference demands low latency, ideally within a single device touchdown. Third, comprehensive control and management are needed for model deployment, updates, selective data ingestion, and the dynamic, distributed test infrastructure, bridging upstream operations and OSAT test floors.
To address these challenges, we developed a scalable architecture integrating cloud and IoT (Internet of Things) technologies. Cloud services, specifically AWS (Amazon Web Services), support centralized data aggregation and ML model training, providing scalable storage and computing resources while enabling fabless teams to develop and maintain models independently of the OSAT environment. For low-latency inference, edge servers are deployed near testers to execute ML workloads locally with small, preprocessed die feature data ingested from cloud. IoT technologies serve as the crucial synchronization layer, enabling real-time bi-directional data and control exchange between cloud and edge. AWS IoT was selected for its seamless cloud integration, fine-grained device control, and secure data exchange capabilities.
The architecture implementation was explored with Advantest ACS Gemini, an AWS-hosted digital twin environment designed for ML-driven test application development. Key components of the solution include: • Aggregating test datalogs into AWS S3, followed by ETL (extract, transform, load) for model training. • Automated provisioning of IoT infrastructure in the production environment including deploying lightweight IoT clients on edge servers and tester host controllers. • IoT-enabled orchestration, including cloud-to-edge commands to deploy containerized models, transmission of selected test-derived die features for inference, and edge-to-cloud alerts for model retraining upon drift detection.
Security, a critical concern in the fabless-OSAT model, is addressed through AWS IoT’s robust identity and policy management, ensuring secure interactions among all participants from initial provisioning of IoT infrastructure through ongoing data exchange.
This presentation will detail the architecture, implementation, practical use cases, findings, and learnings of this solution, demonstrating how cloud-IoT integration enables real-time data-driven optimization in semiconductor test with intelligence, automation, and scalability.
.jpg "Fangmin Chu photo")