Volume diagnosis

Notes

A standard yield-learning tool at advanced nodes, where systematic defects often outweigh random particle defects.

Overview

Volume diagnosis is the practice of running scan diagnosis on the fail data of many failing dies and analysing the results together, so that defect causes shared by many parts stand out from one-off random defects. It turns routine production test failures into evidence about the process and the design, such as a layer, via type or library cell that fails more often than it should. It builds on single-die scan diagnosis, which has its own page.

How it works

Fail logs are collected from many failing dies at wafer sort or final test, and each one is diagnosed. Every diagnosis report lists suspects with their properties, such as the layer, net, cell type and likely defect type. On its own, a single report is often ambiguous, with several suspects that could explain the same failures.

Statistics remove much of that ambiguity. Siemens, for example, describes root cause deconvolution, an unsupervised machine learning method that works only from diagnosis reports and estimates which root causes, such as an open on a particular metal layer, explain the failures and how many devices each accounts for. Engineers then pick the dies most likely to show the cause of interest and send those to physical failure analysis, so that each costly analysis is more likely to confirm something that matters.

Throughput is the practical limit. Diagnosis of a large design takes significant compute, and cell-aware diagnosis, needed for defects inside the transistors of standard cells, makes each run heavier. Techniques such as loading only the part of the design relevant to each fail log cut memory and runtime, and the value of the method grows with the number of dies diagnosed. Results are still bounded by the fault models and the resolution of the underlying diagnosis.

Examples

• GlobalFoundries and Mentor (now Siemens) combining cell-aware volume diagnosis with root cause deconvolution to reduce the number of suspects for defects inside cells.
• Choosing which failing dies to send to physical failure analysis so that each one is likely to show the dominant root cause.
• Tracking the estimated share of each root cause week by week during a ramp, to check whether a process fix has worked.
• Dynamic partitioning in Tessent Diagnosis, which Siemens says typically cuts memory per report about fivefold and raises overall diagnosis throughput about tenfold.

Why it matters

During a yield ramp, knowing which defect cause accounts for the most lost dies tells the fab or the design team what to fix first. Yield enhancement and yield analytics engineers run and interpret volume diagnosis, test data engineers keep fail data flowing from the testers, and failure analysis teams use it to choose samples.

Related terms

• Scan diagnosis
• Wafer sort
• Outlier screening
• SLM

Where it appears on the Hub

Used on 5 pages of the Hub. In each group, the pages that mention it most come first.