Experiences Estimating Test Quality And Test Escape Rates

When discussing product quality in integrated circuits (ICs), two key aspects are essential: time zero defects and reliability. These concepts help distinguish between issues that appear immediately after manufacturing and those that occur over time. Understanding these distinctions is critical for ensuring that products meet the quality demands of their respective markets, particularly in high-stakes sectors like automotive and industrial electronics.

Time zero vs. reliability

Reliability refers to an IC’s ability to function as intended over its operational lifespan. Failures related to reliability are not typically detectable immediately after production but manifest over time. These failures are measured using failures in time (FITs), which assesses the number of failures expected per billion hours of operation.

On the other hand, time zero defects are faults detectable right after manufacturing, under specific conditions. The test escape rate—measured in defective parts per million (DPPM)—captures the probability of these defects going undetected. Both FITs and DPPM are used to model product quality, but the focus here is on DPPM estimation.

Why focus on time zero DPPM estimation?

Different IC market segments, such as automotive or consumer electronics, have varying thresholds for acceptable quality. The higher the quality level required, the more effort and cost it takes to achieve it. For example, the automotive sector demands extremely low DPPM rates due to the safety-critical nature of the applications. By using models to estimate time zero DPPM, companies can match product quality to the various market segments that they serve.

Accurately estimating DPPM is crucial during product development. It allows engineers to make critical design and testing decisions that affect the trade-offs between test effort and product quality. Tracking expected versus actual DPPM during production is a vital process to ensure that any deviations—whether from silicon issues or test execution problems—are caught and addressed early.

Challenges to DPPM fidelity

As IC technology evolves, so too do the challenges of maintaining accurate DPPM models. New defect mechanisms arise, and the way defects cluster may change, adding complexity to the estimation process. Additionally, as IC design and test methods advance, so do the fault models used to generate tests, compute coverage, and estimate DPPM. New techniques, such as adaptive voltage and frequency scaling, introduce more variables into the testing process, making it challenging to model all conditions accurately.

Data streaming and bi-directional test cell control

Effective DPPM estimation relies heavily on empirical data collection. Coverage data from different types of tests—such as Automatic Test Pattern Generation (ATPG)—must be aligned with test content as it is applied in production as well as observed failure rates to refine an empirically-based model. Historically, this was a static exercise, but with the advent of real-time streaming data, monitoring has become a dynamic, continuous process. This shift allows for quicker identification of deviations, which could signal issues with the silicon, test equipment, or other factors. To that end, the Advantest ACS Real-Time Data Infrastructure (ACS RTDI) enables customers to have more immediate access to their test data for many purposes including DPPM estimation, AI/ML analytics in real-time, product and test solution health monitoring, and others. If excursions are detected, ACS RTDI also enables the ability to react quickly to exert control over test operations and perform corrective measures. In total, the new ACS ecosystem enables customers to improve quality, increase yield, enhance operational efficiencies, and ultimately, accelerate product development and new product introductions.

Conclusion

After more than 50 years of DPPM modeling, the field continues to evolve. The foundational mathematical models remain critical, but as IC technology, design, and testing methods advance, so too must the models. Empirically-derived models that can adapt to real-time data and new fault mechanisms are likely the future of DPPM estimation, making this an exciting area for ongoing research and development.