Data Feed Forward And How It Works: Part 1

With data analytics, manufacturers can gain unparalleled insights into their testing processes, identify patterns, predict failures, and optimize operations. From improving yield rates to reducing testing costs, data analytics not only enhance the quality of semiconductor devices but also drives innovation and competitiveness in the industry.

Traditionally, data analytics has been performed “after the fact”: data was collected after the whole lot had been tested, and it was mostly useful to summarize the results, but could not be used to influence and improve the results of the test during the test itself.

Two new data analytics trends are emerging in the industry.

1. Real-time inference during the test process

Cloud-based analysis used to be the only way to crunch massive volumes of test data. But in high-volume production, every second matters. Delays in data transmission and analysis can hurt throughput.

That’s why real-time edge computing is a game-changer.

With ACS RTDI, Advantest’s edge analytics solution, ML inference runs right at the tester. That means analysis is done in milliseconds — not seconds — enabling real-time feedback and adaptive decision-making during test. No roundtrips to the cloud, no bottlenecks — just smart, scalable, and secure defect detection right where it’s needed.

2. Comparing real-time device data across multiple insertions

Use of device test data from previous insertions can help introduce new types of tests that were not possible before:

• Monitor device parameter drifts

• • Devices with changes in parametric response across test/stress can be a quality or reliability risk. Changes can be subtle to detect. Yield can be impacted if using simple deltas or other statistics.
  • Feed forward data from earlier test insertion(s), analyze combined data AI/ML methods to downgrade at-risk devices. Improve quality, reliability, reduce customer returns and associated rework/replacement costs.

• Speed up “search-based” tests

• • Minimum operating voltage (VMIN) searches require long test times.
  • It’s possible to utilize inferencing to predict likely VMIN per-device and narrow the search limits to speed searches.
  • Data Feed Forward on VMIN from previous insertions can help reduce the search limits so that significant test time reduction can be achieved.

• Implement re-binning strategies

• • Conventional pass/fail testing may not suffice to identify devices at risk for early life reliability failures, thus requiring advanced analytics.
  • In multi-tiered binning, complex analytics may be required to upgrade or downgrade devices to the appropriate bin.
  • Re-binning criteria are subject to change frequently, so it is desirable to decouple the analytics from the test program to avoid frequent program releases.
  • With RTDI and DFF, during the testing of a device, it’s possible to modify its binning based on the binning it had in a previous test insertion.

The limits of traditional stats – and the power of machine learning

Simply tightening test limits might seem like a simple fix, but it comes at a cost. Galaxy Semiconductor’s SEMICON West study painted a clear picture: aggressively narrowing dynamic part average testing thresholds caught only 2% of known failures, while scrapping over 12% of good devices.

The better answer? Machine learning.

Both real-time analytics and Data Feed Forward help achieve this.

By applying ML models trained on known failure data, 88% of the same failures were flagged—with just a 2.4% yield hit. ML thrives in complexity, uncovering patterns and subtle anomalies that escape both human intuition and conventional algorithms.