Common Mistakes

Costly Mistakes in Advanced Packaging and Test, and How to Catch Them

The eight errors that quietly wreck back-end yield, capacity, and cost numbers, each with the symptom that gives it away and the fix expressed as a number.

The most expensive back-end mistake is confusing step yield with composite yield. Symptom: your Advanced Packaging Yield calculator shows 99 percent per operation, yet finished modules come out near 94 percent and nobody can explain the gap. Root cause: yields multiply, they do not average. Six operations at 0.99 give 0.99 to the sixth power, which is 0.941, not 0.99. Fix: always stack step yields as a product. On a chiplet module with eight bonds at 0.995 each, composite lands at 0.961, so a single step slipping to 0.98 drags the module to about 0.945, roughly 1.6 points of finished loss from one step.

The second error is quoting gross capacity as committed capacity. Symptom: planning promises 1,920 good die from a probe cell, and the line delivers around 1,676. Root cause: gross output of 4 die per cycle times 480 cycles ignores the uptime and first-pass yield derates. At 90 percent uptime and 97 percent yield, Wafer Sort Capacity returns 1,676, with 192 units lost to downtime and about 52 to yield. Fix: never commit the raw product. Multiply gross by both derates, and the same trap sits inside Final Test Throughput and Burn-in Capacity, so audit all three the same way.

Third is treating burn-in yield fallout as a defect. Symptom: post-burn-in first-pass yield sits at 97 percent and a manager launches a corrective action to recover the missing 3 percent. Root cause: burn-in exists to precipitate infant-mortality failures, so some fallout is the screen working as designed, not a process excursion. Fix: separate expected survival fallout from genuine defect fallout using historical baselines. If a stable automotive part normally loses 2 to 4 percent at burn-in, a lot at 3 percent is healthy, while a lot at 8 percent signals a real weak-population problem worth chasing, not the baseline 3 percent.

Fourth is a units mistake on test cost: mixing cost per insertion with cost per second. Symptom: Package Test Cost per good device comes out roughly ten times too high or too low. Root cause: entering a full per-insertion tester charge into a per-second field, or vice versa, when a 3 second insertion at a 0.0007 dollar per second machine rate is about 0.0021 dollars, not 0.0007. Fix: pin the tester rate to one basis. If your ATE amortized rate is 90 dollars per hour, that is 0.025 dollars per second, so a 4 second insertion is 0.10 dollars before yield and setup are folded in.

Fifth is forgetting to divide test and yield cost by good units instead of tested units. Symptom: cost per unit looks fine until a low-yield lot ships and margin evaporates. Root cause: dividing total test spend by devices inserted rather than devices that passed. At 92 percent first-pass yield, cost per good device is total cost divided by 0.92, an 8.7 percent uplift over the naive per-tested figure. Fix: always denominate cost per good unit, and re-run Package Test Cost whenever yield moves more than 2 points, since the per-good number swings faster than the headline yield.

Sixth is undervaluing scrapped assemblies by counting units, not dollars. Symptom: two lines report the same 3 percent scrap and get equal attention. Root cause: a scrapped 2.5D module carrying four known-good die plus an interposer can hold 40 times the value of a scrapped single-die package, so unit-based scrap hides where money actually burns. Fix: weight yield loss by value at risk. Run Chiplet Assembly Yield and Flip Chip Yield alongside the die value, so a 3 percent loss on modules worth 300 dollars each is a 9 dollar per module exposure, dwarfing an equal percentage on 8 dollar parts.

Seventh is amortizing probe cards and load boards over the wrong denominator. Symptom: cost per die on a high-pin-count card looks cheap on paper, then blows the budget. Root cause: spreading card price over nameplate lifetime instead of realized touchdowns. A 45,000 dollar probe card rated for 1 million touchdowns but retired at 600,000 costs 0.075 dollars per touchdown, not 0.045. Fix: amortize on actual touchdown-limited life and cleaning intervals, and feed that into Substrate Cost and Package Test Cost so tooling and NRE land on the real build volume, not an optimistic one.

Eighth is reading small-lot yield as a stable process number. Symptom: a bring-up lot of 250 assemblies with 8 good reports 3.2 percent, and someone extrapolates it as the line entitlement. Root cause: at high target yields a single unit can swing the percentage by 0.4 points, and 250 units is far too few to bound a 95 percent process. Fix: require a few hundred to a few thousand units before trusting a yield read, and pair Die Attach Yield and Wire Bond Workload with a defect Pareto so you attack the dominant failure mode instead of chasing statistical noise on tiny lots.

Published 2026-07-01.