
A die that passes sort is not a good die. It is a die that is probably good, and the size of that "probably" depends on its yield and on how much of it the test could see. In the illustrative six-die package modelled below, a small die worth 0.6% of the bill of materials accounts for 3.0 of the 7.9 percentage points of package yield loss, because each of its escapes scraps everything bonded around it.
That is the economic case for looking past the words "known good". This post builds the yield stack one line at a time, tests three ways of spending money on the problem, and then looks at what IEEE Std 1838 does and does not contribute once dies are buried in a stack.
Key takeaways
- Package yield multiplies, so the die that matters most is the one with the worst outgoing quality, whatever it costs to make.
- In our illustrative model, 76 cents more spent on a $7 die cuts the cost of a good package by $29.83. Those figures are assumptions worked through, not industry data.
- Published cost studies warn against rules of thumb: the best test flow depends on each product's parameters.
- IEEE Std 1838 standardises how test data reaches each die in a stack. It does not decide your coverage, your fault models or your test flow.
- Field data suggests escapes are already a live problem for compute chips. A Google study reports about 5,000 defective parts per million over lifetime against an expected 100 to 500 or less.
Known good is a probability
Wafer sort removes the dies it can prove are bad. Some bad dies pass, because no test covers every fault. A common way to estimate how many is the Williams and Brown model, which Agrawal and Chakrabarty use in their test-flow study at Duke University. If Y is the true yield of the die and T is the fault coverage of the test:
- The fraction of dies that pass the test is YT.
- The fraction of passing dies that are in fact defective, the defect level, is 1 minus Y(1 minus T).
The model is a simplification, but it shows that yield and coverage act together. A die with 85% yield and 80% coverage ships with a defect level of 3.2%. The same die at 95% coverage ships at 0.81%.
In a single-die product, an escape costs one die, one package and some test time. In a multi-die product it costs every other die in the package as well. A November 2024 Open Compute Project paper on system-in-package testing, written by authors from test and design companies, makes the same point: one bad die can fail the whole package.
A worked yield stack
Every number in this section is an illustrative assumption, not data from any foundry, vendor or product. The package holds six dies: two large compute dies, one I/O die, two bought-in memory stacks, and one small mixed-signal die whose fine-pitch bumps and analogue content leave it with weak structural coverage at sort.
| Die | Qty | Cost before sort | Sort test | Yield Y | Coverage T | Pass rate | Cost per passing die | Defect level |
|---|---|---|---|---|---|---|---|---|
| Compute | 2 | $120 | $3.00 | 70% | 99% | 70.25% | $175.09 | 0.36% |
| I/O | 1 | $30 | $1.50 | 90% | 98% | 90.19% | $34.93 | 0.21% |
| Mixed-signal | 1 | $6 | $0.25 | 85% | 80% | 87.81% | $7.12 | 3.20% |
| Memory stack | 2 | Bought tested at $300 each, with an assumed 99.5% good rate | $300.00 | 0.50% | ||||
Cost per passing die is the cost before sort plus the sort test, divided by the pass rate. Add a $100 substrate, $60 of assembly and $20 of final test, and each package built costs $1,172.22. Assembly yields 99.5% per attach, across six attaches. Totals carry unrounded values, so a sum of rounded lines can differ by a cent.
Now stack the yields, assuming independent failures, no repair and a final test that catches every failure:
| Contributor | Multiplier | Running package yield |
|---|---|---|
| Compute dies (2) | 99.29% | 99.29% |
| I/O die | 99.79% | 99.08% |
| Mixed-signal die | 96.80% | 95.91% |
| Memory stacks (2) | 99.00% | 94.95% |
| Assembly (6 attaches) | 97.04% | 92.14% |
Package yield is 92.14%, a loss of 7.9 points. If the mixed-signal die never escaped, yield would be 95.19%, so that one die accounts for 3.0 of those points. Removing all six assembly losses recovers 2.8. No other line comes close, including the two compute dies, which together cost nearly 50 times as much.
The die is 0.6% of the bill of materials, at $7.12 out of $1,172.22. Each escape, however, throws away the whole $1,172.22, about 165 times what the die itself cost. Dividing the build cost by the yield gives $1,272.20 per good package, of which about $100 is scrap.
Three ways to spend on it
Here are the options, using the same illustrative model. A mid-stack test is a $4 test on the part-built package, before the memory stacks go on. We assume it catches 80% of failing partial assemblies with no false rejects, that $40 of the assembly cost falls before it, and that the partial assembly is 94.01% good at that point.
| Scenario | Package yield | Average spend per build | Cost per good package | Saving against baseline |
|---|---|---|---|---|
| Baseline: mixed-signal coverage 80%, $0.25 test | 92.14% | $1,172.22 | $1,272.20 | None |
| Better sort: 95% coverage, $0.75 test | 94.42% | $1,172.98 | $1,242.36 | $29.83 |
| Best sort: 98% coverage, $1.00 test | 94.88% | $1,173.31 | $1,236.67 | $35.53 |
| Baseline plus mid-stack test | 92.14% | $1,145.54 | $1,243.24 | $28.96 |
| Better sort plus mid-stack test | 94.42% | $1,158.18 | $1,226.68 | $45.51 |
- The cheap fix is on the cheap die. Moving to 95% coverage costs 76 cents per package once the extra sort rejects are counted. The cost per passing die goes from $7.12 to $7.88. Break-even for that yield gain is $28.93 of added spend per package, so the margin for error is wide.
- Returns fall quickly. Going on to 98% adds only 0.46 points of yield and $5.69 of saving, because the mixed-signal die has stopped being the worst line in the stack. Assembly now is.
- Sort and mid-stack test do different jobs. Mid-stack test leaves yield at 92.14% and saves money by not bonding $600 of memory onto assemblies that are already bad. Better sort keeps a bad die out in the first place. Together they save $45.51, which is less than the sum of the two, since each one reduces what the other can catch.
Final test matters too. Using the same model, a 99% coverage final test on a population that is 92.14% good lets through about 820 defective parts per million. At 94.42% good, that drops to about 575. Better incoming quality reduces escapes as well as scrap.
What the published studies say
No universal flow exists. Agrawal and Chakrabarty searched the possible test flows for 3D stacks instead of hand-picking a few. A four-die example had 8,192 possible flows with one test per insertion, and the worst flow cut profit margin by 35% against the best. Their conclusion is that optimal choices depend on the problem instance, so a generic set of rules cannot be given.
Dies higher in the stack justify more pre-bond test. In that same example, with one test of 100% coverage available and all other die yields at 0.99, pre-bond test stayed in the optimal flow up to a die yield threshold of 0.88 for the first die, 0.94 for the second, 0.96 for the third and 0.98 for the fourth. A defective die bonded onto a stack already known to be good wastes all of it. Their parameters are assumptions too, so read the trend and not the values.
Interposers and mid-bond steps can swing the total. TU Delft and imec built the 3D-COSTAR cost tool, which models pre-bond, mid-bond, post-bond and final test for 2.5D and 3D stacks. Its project page quotes imec's Erik Jan Marinissen saying that pre-bond interposer test reduced overall stack cost by up to 40% in some cases. The same page says the tool showed when mid-bond test could be skipped without much effect on stack yield. It also carries a vendor claim from Cascade Microtech of up to 50% overall cost benefit from direct microbump probing. Those figures date from around 2013 and depend on the case studied.
The assembly term deserves a real-world check as well. In a March 2023 IMAPS presentation, Adeia reported a 12-die hybrid bonded test stack with 56% overall yield and an average of 95% per layer. Multiplying 95% twelve times gives 54%, close to the reported figure once rounding in the average is allowed for. This is vendor material on a test vehicle, and the company notes that an early particle excursion depressed yield before per-layer results improved. Our post on hybrid bonding from first principles covers the bond itself.
Then there is what happens after shipping. A 2025 preprint by Mitra and colleagues, drawing on Google's fleet, reports test escapes of around 5,000 defective parts per million over lifetime across compute chip types, against an industry expectation of 100 to 500 or less. About 1,000 of those are estimated to cause silent data corruption. That is one operator's lifetime data on compute chips in general, not multi-die parts, so it should not be set against the shipment estimates in our model.
Where IEEE Std 1838 fits
All of this assumes you can reach the die. A 2016 International Test Conference overview of the then-draft standard, given by Adam Cron of Synopsys, listed the access problems at each step: TSVs that are not accessible, contacts too small and too close together, and contacts that disappear once bonded.
IEEE Std 1838-2019, the Standard for Test Access Architecture for Three-Dimensional Stacked Integrated Circuits, is the formal response. The IEEE lists it as active, with board approval on 7 November 2019 and publication on 13 March 2020. Its main features, per the IEEE page and an imec release of January 2020, are below. The last column is our reading.
| Element | What it does | Why the economics care |
|---|---|---|
| Die-centric compliance | Applies to a die intended for a stack, with die-level features that combine into a stack-level architecture | Dies from different makers can share a test route, if each maker built to it |
| Die wrapper register (DWR) | Scan chains at each die boundary for modular test of die internals and of interconnect between adjacent dies | Lets a failure be assigned to a die or to a bond, which yield learning depends on |
| Serial control mechanism (SCM) | A single-bit channel carrying instructions into the stack to set wrapper test modes | Gives low-cost control through the stack at any stage |
| Flexible parallel port (FPP) | An optional, scalable multi-bit path for moving large test data up and down the stack | Test time after stacking, and so the cost of any mid-stack test, depends on this bandwidth |
The IEEE scope covers test of intra-die circuitry and inter-die interconnect before and after stacking, on partial and complete stacks, at pre-packaging, post-packaging and board level. Those are the insertions a cost model chooses between. The main focus is through-silicon vias, though other interconnects such as wire bonds are not excluded.
The limits matter just as much. The working group's introduction says dies may have their own technologies and test approaches and that the standard should not modify them. So, as we read it, 1838 is a delivery route. The coverage figure in our table above, which drove the whole result, stays with the die maker's own DFT and test program. A compliant die can still have 80% coverage.
Two related efforts extend this to other package types. The IEEE page lists a proposed amendment, P1838a, in development when we read it on 7 November 2026, for testing external interconnects of multi-die assemblies such as 2D and 2.5D packages using boundary-scan segments. Separately, IEEE P3405, Standard for Chiplet Interconnect Test and Repair, had its project authorisation approved on 21 September 2023 and was still listed as an active project when we checked on 7 November 2026. It covers clustering, clocking, spares, repair multiplexing, lane numbering and repair signatures.
Repair changes the sum
Our model has no repair, which makes it pessimistic for some interconnects. A 2023 Synopsys article, which is vendor material, describes UCIe test and repair running at production test and at link initialisation, rerouting data to predefined spare lanes, with up to 8 spare pins per direction in advanced package configurations. If repair recovers most lane defects, the assembly multiplier moves closer to 100%, and the balance shifts further towards die quality. Our post on UCIe in practice looks at what the specification leaves to implementers.
Building your own test plan
Ask these while the package is being partitioned and the DFT can still change. Our chiplet and advanced packaging lifecycle on The Semiconductor Hub shows where these decisions sit in a programme.
- What is the defect level of each die, not just its yield? Estimate both yield and coverage for every die, including the small ones and any bought-in part.
- Which line is worst in the yield stack? Rank by multiplier. Cost per die is the wrong ranking.
- What is the break-even spend? Divide new yield by old yield, subtract 1, and multiply by the build cost. Our example gives $28.93.
- In what order are dies bonded? Put the expensive parts as late as the process allows, with a test point before them.
- Can each insertion actually reach the die? Confirm that sort access, wrapper modes and stack bandwidth exist for each test the plan assumes, before tape-out.
Sources
- IEEE Standards Association, IEEE 1838-2019 project page, "IEEE Standard for Test Access Architecture for Three-Dimensional Stacked Integrated Circuits", approved 7 November 2019, published 13 March 2020, including the P1838a listing. Read 7 November 2026.
- IEEE 3D-Test Working Group page, grouper.ieee.org/groups/3Dtest. Read 7 November 2026.
- Cron (Synopsys), "3D-IC Test Standard IEEE P1838: Overview", International Test Conference 2016 special session slides, 15 November 2016.
- imec, release on the publication of IEEE Std 1838-2019, 27 January 2020, read as reprinted by EE Journal.
- IEEE Standards Association, P3405 project page, "Standard for Chiplet Interconnect Test and Repair", PAR approved 21 September 2023. Read 7 November 2026.
- Agrawal and Chakrabarty (Duke University), "Test-Cost Modeling and Optimal Test-Flow Selection of 3D-Stacked ICs", report extending their VLSI Test Symposium 2013 paper. It applies the Williams and Brown defect level model of 1981.
- TU Delft, 3D-COSTAR tool page, with imec and Cascade Microtech statements from the 2013 release. Read 7 November 2026.
- Mitra et al. (Google), "Silent Data Corruption by 10x Test Escapes Threatens Reliable Computing", arXiv:2508.01786, revised 21 August 2025.
- Bartley, Wong, Ang and Sanyal, "System in Package Testing", Open Compute Project paper, November 2024.
- Workman et al. (Adeia), "Hybrid Bond Interconnect and Advanced Packaging Solutions", IMAPS Conference on Device Packaging, March 2023. Vendor material.
- Mota (Synopsys), "Leveraging UCIe IP for Multi-Die System Reliability", 17 April 2023. Vendor material.
The worked model uses illustrative numbers throughout and is no guide to any real product's yield or cost. Vendor figures are the companies' own claims, as published on the dates shown.


Your comments will be moderated before it appears here.