HomeChiplets & UCIeDay 8
DAY 8 · PHASE 2 — ADVANCED PACKAGING

Known Good Die (KGD)
Yield Multiplication & Test Economics

By EcrioniX · Updated July 2026

Days 6 and 7 covered how chiplets get bonded together. Day 8 covers the question that determines whether any of it is worth doing: how do you know a chiplet works before you've permanently bonded it to three or eleven others? Once hybrid-bonded, there is no rework. This lesson is about the economics and engineering of testing chiplets before that point of no return.

The Point of No Return

A monolithic die that fails test is scrapped — one wasted die. A multi-chiplet package is a different story entirely: once chiplets are joined via micro-bumps or hybrid bonding, rework is impossible. If even one of several bonded chiplets turns out to be defective, the entire package — every good chiplet along with the bad one — gets scrapped. A hybrid-bonded stack in particular cannot be reheated and separated without destroying it.

The cost of catching a defect climbs steeply the later you catch it. A defect found at wafer-level die test is cheap to deal with — discard one die. The same defect discovered only after full packaging can cost roughly 10 to 100 times more, because it now takes the entire assembly, the bonding process, and every other good chiplet in that package down with it.

Yield Multiplication — The Math That Forces the Issue

System yield across N independently-tested chiplets is approximately the product of each chiplet's individual yield. This compounds fast:

Worked example — composite yield
4-chiplet package, 95% known-good-die yield per chiplet:
  0.95 ^ 4  =  0.8145   →   ~81.5% composite yield

AMD MI300X-scale package, 12 chiplets, 98% known-good-die yield per chiplet:
  0.98 ^ 12 =  0.7847   →   ~78.5% composite yield

Even at a high 98% individual chiplet yield, stacking 12 of them
still throws away roughly 1 in 5 completed packages.

This is the sharpest possible argument for rigorous pre-bond testing: the more chiplets you combine — exactly the strategy this course has spent seven days explaining the benefits of — the more brutally individual chiplet defect rates compound against you at the package level, unless you catch bad die before they're ever bonded.

KGD Test Economics — Why It's Worth Paying More

💵
$5–$15 per chipletTypical cost of thorough structural, parametric, and BIST testing at the die level
📈
10–100x cost multiplierHow much more expensive it is to catch the same defect after packaging instead of at the die
Up to 16x justifiedKGD testing can cost up to 16 times more than testing an unrepairable single-die module and still be the economically correct choice

In practice, testing every die to the fullest possible standard isn't always economically feasible either — which is why the industry compromise term "Good Enough Die" exists: balancing test thoroughness and cost against the real fallout risk for a given product's chiplet count and value.

The Three Test Stages

StageWhenWhat It Checks
Pre-bond wafer sort (EWS)Before any assemblyElectrical Wafer Sort tests each die on the wafer, sorting known-good die before they're ever diced and bonded
Known-Good-Stack (KGS) testAfter 3D stacking, before final packagingVerifies the bonded stack as a unit — catching bonding-process defects that individual pre-bond tests couldn't see
Final (post-package) testAfter complete assemblyValidates the fully packaged part before it ships to a customer

Why You Can't Just Probe It Normally

Testing at fine bump or hybrid-bond pitch runs into a physical wall: the pads are simply too small and too closely spaced for conventional test probes to touch directly. This is exactly the same class of problem as any other design-for-test challenge — you need built-in structures (scan chains, BIST, boundary-scan-style access) designed into the chiplet from the start, because you cannot rely on external probing once the pitch gets fine enough.

IEEE P1838 is the industry's answer at the standards level: a test access architecture standard for 2.5D, 3D, and 5.5D stacked integrated circuits, defining how test signals reach into chiplets whose physical connections can't be probed directly, so pre-bond, post-bond, and stack-level testing can all be coordinated through a common, standardized access scheme rather than a different bespoke solution per product.

🎯 Day 8 Key Takeaways

Frequently Asked Questions

What is Known Good Die (KGD) testing?
Known Good Die testing is thorough structural, parametric, and built-in-self-test (BIST) validation of an individual chiplet before it is bonded into a package, to confirm the die is fully functional. It matters far more for chiplets than for a single monolithic die because, once bonded via microbumps or hybrid bonding, rework is impossible.
Why does yield multiplication make KGD testing so important?
System yield across N chiplets is roughly the product of each chiplet's individual yield. For a 4-chiplet package at 95% known-good-die yield each, composite yield before assembly losses is only 0.95 to the 4th power, about 81.5%. AMD's 12-chiplet MI300X, even at 98% per-chiplet yield, still lands at only 0.98 to the 12th power, about 78.5% composite yield.
How much does KGD testing cost?
Thorough KGD testing — structural, parametric, and BIST — typically costs $5 to $15 per chiplet. Because a single undetected defect can scrap an entire multi-chiplet package once bonded, KGD testing can cost up to 16 times more than testing an unrepairable single-die module and still be economically justified.
What are the test stages for chiplets before and after bonding?
Chiplets are first tested pre-bond at the wafer level (electrical wafer sort, EWS) to sort known good die before assembly. After 3D stacking, a known-good-stack (KGS) test verifies the bonded stack. A final post-package test then validates the complete assembled part before it ships.
What is IEEE P1838?
IEEE P1838 is the standard for test access architecture for 2.5D, 3D, and 5.5D stacked integrated circuits, defining how test signals reach chiplets whose fine-pitch micro-bumps or hybrid-bond pads cannot be directly probed with conventional test equipment.