Diagnosis & Yield Learning: Theory Deep Dive

Foundational theory

Diagnosis & Yield Learning is central to Compression & Diagnosis. Fail logs and diagnosis map failing signatures back to likely defect sites, enabling systematic yield learning loops with design and process teams. Senior DFT engineers tie metric movement to architecture assumptions, constraints, and silicon evidence rather than isolated tool output.

Core concepts explained

• Fail logs and diagnosis map failing signatures back to likely defect sites, enabling systematic yield learning loops with design and process teams.

• Primary metric: diagnosis resolution, top failing buckets, yield learning turnaround

• Primary artifact: diagnosis report, fail-log bucket table, yield pareto dashboard

• Owners: product test owner, yield engineer, DFT lead

• Controllability and observability must be explicit

• Production-quality requires reproducible pattern and tester tags

Why this matters at release

At release, Diagnosis & Yield Learning issues can create coverage escapes, unstable production bins, or long debug loops. Compression is a quality-and-cost optimization, not just a ratio target.

Mental model

fail logs -> suspect list -> diagnosis buckets -> yield pareto -> design/process action

The loop is valuable only with consistent signatures and tagging.

Worked intuition

1. Name failing metric and scenario context (mode, lot/corner, program).

2. Open diagnosis resolution, top failing buckets, yield learning turnaround trend and isolate dominant failing bucket.

3. Trace architecture assumptions and legality constraints.

4. Check compression, clocking, and unknown handling dependencies.

5. Collect diagnosis report, fail-log bucket table, yield pareto dashboard and confirm run tags.

6. Classify issue: model/constraint, physical/test setup, or real defect signal.

7. Propose minimal fix and list timing/power/quality regression checks.

Common misconceptions

• Coverage percent alone proves release readiness.

• More compression always means better outcome.

• Silicon mismatch can be debugged without pattern/tester traceability.

• Shift timing and test power can be signed independently.

Visual reinforcement

Diagnosis feedback loop

fail logs -> suspect list -> diagnosis buckets -> yield pareto -> design/process action

The loop is valuable only with consistent signatures and tagging.

Layer responsibilities

DFT OWNERSHIP LAYERS - Diagnosis & Yield Learning

layer              owns                         failure mode
----------------   --------------------------   -------------------------
rtl/architecture   scanability hooks            uncontrollable logic
atpg/constraints   legal pattern intent         aborts, low coverage
physical/clocking  chain route + test clocks    shift hold/timing escapes
tester/program     pattern apply integrity      false binning / bad fails
quality signoff    release criteria             escapes or schedule slip

DFT deep dive

Compression saves tester time only when diagnosis observability remains credible.

Concept diagram

COMPRESSION LOOP

EDT/decompressor -> compressed patterns -> compactor responses -> diagnosis

Metric graph

PATTERN vs COVERAGE

coverage up   -> pattern count up
compression up -> pattern count down (until aliasing risk)

Reports and artifacts

• compression ratio dashboard

• pattern count trend

• X-source report

• diagnosis bucket summary

Mini case study

Compactor overflow plus unknown inflation caused false diagnosis; masking policy and channel map corrected.

Debug branches

• Separate X issues from silicon defects

• Replay failing patterns uncompressed

• Track tester memory budget

Senior review question

Ask: what evidence proves this DFT decision is safe for production?

Key takeaways

• State metric, lot/corner context, and pattern tag with every claim.

• Treat timing, power, and quality as one signoff problem.

Common pitfalls

• Chasing coverage without legality checks.

• Ignoring test-power side effects of pattern changes.

• Debugging silicon without reproducible tags.

Theory reinforcement

Compression is a quality-and-cost optimization, not just a ratio target.