U-Series: Cross-Period NGBin 19/109 Yield Excursion and Wafer-Edge Analysis

Overview

U-Series repeatedly showed abnormal NGBin 19 Full Erase Fail and NGBin 109 Full Block Read Fail during MBT across three separated excursion windows from July 2020 through March 2021.

The analysis separated the monthly history from a March weekly drill-down. W11 was treated as an early signal, while W12–W13 formed the third excursion. April was treated as the transition month, May–July as the stable post-action window, and August as continued monitoring outside the primary comparison.

The investigation combined MBT retest review, backend equipment checks, internal trace data, Spotfire yield and bin visualization, and normalized wafer-map analysis. Across the consolidated excursion windows, the outermost three complete die rings showed a 5.77× NGBin 19/109 rate relative to the interior after normalization by tested-die population.

Fab teams applied generalized CVD- and TEOS-related controls to reduce upstream defect generation, while EDS teams adjusted wafer-level screening to reduce downstream escape. The combined-package rate declined from 3,250 ppm in March to 243 ppm across May–July, an absolute rate reduction of 3,007 ppm.

My direct contribution included consolidating monthly and weekly yield and NGBin trends; defining excursion and observation windows; reviewing MBT retest behavior and backend equipment concentration; tracing relevant records through internal manufacturing systems; analyzing wafer-map patterns in Spotfire; consolidating repeated excursions to identify the wafer-edge tendency; escalating the evidence to fab and EDS teams; monitoring downstream performance; and constructing the cost-impact model. Fab occurrence controls and EDS screening changes were implemented by the respective cross-functional teams.

Terminology

• UFS: Universal Flash Storage, an embedded flash storage product type.
• U-Series: An anonymized UFS product family used in this case study.
• MBT: An anonymized backend memory-test stage used in this case to screen NAND-related failure behavior.
• NGBin: An internal failure-bin category used to classify test failures.
• NGBin 19: Full Erase Fail.
• NGBin 109: Full Block Read Fail.
• EDS: Electrical Die Sorting, the wafer-level electrical test used to screen dies before backend assembly.
• Wafer map: A spatial map of chip-level results across a wafer, used to identify location-based yield patterns such as center, edge, ring, or random distributions.
• CVD: Chemical Vapor Deposition, a fab process used to deposit thin-film layers.
• TEOS: Tetraethyl orthosilicate, a silicon oxide precursor used in semiconductor deposition processes.
• Spotfire: Visual analytics software used here for yield, bin, wafer, and defect-pattern analysis.

Problem

The same NGBin 19 and NGBin 109 categories recurred across separated windows, triggering abnormal handling and WIP holds. Single-window wafer checks gave no stable explanation, and no clear backend equipment concentration was identified in the reviewed early windows.

The investigation had to distinguish backend test noise or equipment concentration from an upstream spatial tendency or combined occurrence and screening mechanisms.

Data Used

• Monthly U-Series input, yield, and NGBin 19/109 trend data from July 2020 through August 2021
• March 2021 W10–W13 weekly drill-down
• MBT retest behavior used to distinguish persistent defects from simple false-fail noise
• Backend equipment concentration checks
• Internal manufacturing-database trace records
• Spotfire yield, bin, wafer, and wafer-map visualizations
• Normalized edge-versus-interior tested-die populations
• Fab CVD / TEOS feedback and EDS screening-change records at a generalized process-control level
• May–July post-action rate comparison and cost-impact model

Constraints

• Monthly March values and W10–W13 values describe the same period at different aggregation levels and must not be interpreted as consecutive, non-overlapping periods.
• The wafer-edge region is operationally defined for this analysis as the outermost three complete die rings, excluding partial edge dies and non-testable locations; this is not a universal semiconductor-industry definition.
• Normalized edge concentration supports a wafer-level process hypothesis but does not prove that every NGBin 19/109 failure had an identical physical mechanism.
• Fab CVD / TEOS changes address occurrence risk, while EDS changes address screening; the combined post-action comparison does not isolate each action's independent effect.
• The ¥3.48M value is an illustrative modeled cost impact based on stated volume and per-unit-loss assumptions, not a finance-audited realized saving.
• The exact confidence intervals and Fisher tests treat units or dies as independent observations. Wafer- and lot-level clustering was not modeled, so the inferential statistics are descriptive rather than production-grade causal validation.
• No raw production records, wafer IDs, internal recipe values, customer identifiers, or proprietary source documents are published.

Approach

The investigation moved from repeated MBT symptoms through retest, equipment, trace, and normalized wafer-map evidence, identifying a cross-period tendency and separate occurrence and screening action paths.

1. Consolidate repeated monthly excursions rather than treating each month independently.
2. Separate excursion, observation, transition, and stable monitoring windows.
3. Review MBT retest behavior to determine whether failures persist.
4. Check backend equipment concentration across the reviewed early windows.
5. Trace relevant lots and wafer records through internal manufacturing systems.
6. Compare wafer maps across combined excursion windows.
7. Normalize edge and interior failure rates by tested-die population.
8. Escalate the normalized wafer-edge pattern to fab and EDS teams.
9. Separate fab occurrence controls from EDS screening changes.
10. Track the downstream trend, statistical comparison, and modeled cost impact.

Highlights

• Three separated excursion windows were consolidated into one cross-period investigation.
• Monthly history and the March weekly drill-down are displayed as separate aggregation levels.
• Retest, backend-equipment, and internal trace review supported escalation beyond backend testing.
• The normalized edge-to-interior NGBin 19/109 rate ratio was 5.77×.
• The post-action rate declined from 3,250 ppm in March to 243 ppm across May–July.
• Fab occurrence controls and EDS screening improvements were documented as separate action paths.

Key Investigation Choices

Key Findings

The reviewed early windows did not show a stable backend-equipment concentration. In the observation wafer population, the normalized edge-to-interior rate ratio was 1.22× and was not statistically significant in a two-sided Fisher's exact test (p = 0.099, α = 0.05). This supports consolidating repeated excursion windows rather than inferring an edge mechanism from each period independently.

Across consolidated excursion windows, the outermost three complete die rings represented 20.9% of tested dies but 60.4% of NGBin 19/109 failures. The normalized edge rate was 5.77× the interior rate (p < 0.001, two-sided Fisher's exact test). This operational rule is case-specific, not universal.

The action path remained split: fab teams applied generalized CVD / TEOS controls to reduce defect generation, while EDS teams adjusted wafer-level screening to reduce downstream escape.

The post-action rate was lower than the March rate in the combined-package comparison (p < 0.001, two-sided Fisher's exact test). The pattern supports the documented downstream improvement but does not isolate the effect of each individual action.

Corrective Actions

• Fab teams reviewed the repeated wafer-edge pattern and implemented generalized CVD-related interlock timing controls intended to reduce occurrence risk.
• Fab teams implemented generalized TEOS-related process delay/thickness controls intended to reduce occurrence risk.
• EDS teams reviewed wafer-level electrical-test conditions and adjusted upstream screening intended to reduce downstream escape.
• Wafer-edge-pattern monitoring was added to post-change review.
• NGBin 19/109 ppm, weekly recurrence, and downstream MBT behavior were monitored after the combined changes.
• The post-action review used May–July as the stable window and treated April as transition.

Methods & Tools

• Spotfire
• Internal Manufacturing Database
• Trace Analysis Tools
• Wafer Map Visualization

Result & Impact

The combined NGBin 19/109 rate declined from 14,703 failures among 4,524,000 units in March 2021—3,250 ppm, with a two-sided 95% Clopper–Pearson exact binomial confidence interval of 3,198–3,303 ppm—to 3,434 failures among 14,130,000 units from May through July—243 ppm, with a corresponding interval of 235–251 ppm. The absolute rate difference was 3,007 ppm. A two-sided Fisher's exact test showed a difference between the pre- and post-action rates (p < 0.001, α = 0.05).

Within the excursion-window wafer-map population, the outermost three complete die rings represented 20.9% of tested dies but 60.4% of NGBin 19/109 failures. The normalized rate was 13,409 ppm at the edge versus 2,324 ppm in the interior, a 5.77× rate ratio (p < 0.001, two-sided Fisher's exact test).

The observation population showed a smaller 1.22× rate ratio. That difference was not statistically significant in a two-sided Fisher's exact test (p = 0.099, α = 0.05), supporting the decision to combine repeated excursion windows rather than infer an edge mechanism from every period.

Using a representative monthly input of 4.71M units, the 3,006.971 ppm rate reduction is equivalent to an estimated monthly reduction of approximately 14,163 affected units. At an assumed manufacturing loss of ¥246 per affected unit, the illustrative modeled cost impact is approximately ¥3.48M per month.

The comparison evaluates the combined fab occurrence controls and EDS screening changes. It does not isolate the effect of CVD, TEOS, and EDS changes individually or represent a finance-audited realized saving. The exact tests are descriptive and do not account for wafer- or lot-level clustering.

Learnings

• Repeated yield excursions should be consolidated across time before concluding that each window is independent.
• Monthly summaries and weekly drill-downs must be displayed as different aggregation levels rather than consecutive periods.
• Wafer-edge concentration should be normalized by tested-die population; raw spatial counts alone can mislead.
• Retest and backend-equipment review help determine whether an excursion is persistent yield loss or screening noise.
• Fab occurrence controls and EDS screening changes address different parts of the quality chain.
• A combined post-action comparison can document improvement without isolating each action's independent causal effect.
• Public engineering cases should keep analytical assumptions, aggregation levels, and causal boundaries explicit.
• Cost-impact estimates should show production-volume and per-unit-loss assumptions.