Die Yield and Cost per Good Die: Practical Tips and Common Mistakes
Reviewed by the OnlineFree.app team · Updated
Key points
- Cost per good die equals wafer cost divided by the dies that survive yield, not the dies you print.
- The 858 mm² reticle field is a hard manufacturing limit; larger designs need stitching or partitioning.
- Yield falls off exponentially with area, so defect density punishes large dies far more than small ones.
- At 0.1 defects/cm², eight 100 mm² chiplets yield about 91% each versus roughly 47% for one 800 mm² die.
- Chiplet savings are silicon-only; packaging, assembly yield and test must be added before comparing totals.
What does one good die actually cost?
Cost per good die is the price of a processed wafer divided by the number of dies on it that survive yield screening. With the Die Yield And Cost Calculator's defaults — 800 mm² of silicon, 0.1 defects/cm² and a $17,000 300 mm wafer — the tool returns roughly 47% yield, about 62 gross dies, about 29 good dies, and a cost near $590 per good die.
The value of that number is comparison. A yield percentage tells you how well a process is running; cost per good die tells you what you are actually buying. Two designs can both show 70% yield and still differ sharply in cost per die, because one packs far more candidates onto the same wafer.
We built the Die Yield And Cost Calculator to put that figure on one screen next to a monolithic-versus-chiplet comparison, because once someone sees cost per good die, the follow-up question is almost always whether splitting the design would be cheaper.
The 858 mm² reticle limit is a hard wall
A single exposure covers 26 × 33 mm, or 858 mm². Anything larger has to be printed in pieces and stitched, which is why so many large designs land just under that boundary rather than at a rounder number higher up.
The calculator accepts areas up to 3000 mm² on purpose, because people do model stitched or partitioned designs. But past 858 mm² the yield figure is the least trustworthy part of the output: stitching adds alignment and defect risk that a single-density Murphy model cannot see.
The arithmetic still shows why the wall exists. Run 3000 mm² at 0.1 defects/cm² and $17,000 per wafer and you get roughly 10% yield, about 11 gross dies and therefore a single good die — at $17,000 each. That is the shape of the problem the reticle limit protects you from.
Defect density is per cm², not per mm²
Murphy's model, Y = ((1 − e^(−A·D0)) / (A·D0))², wants area in cm² and D0 in defects per cm². Typing 800 where the formula expects 8 inflates A·D0 by 100× and drives yield to nearly zero; it is the most common way a hand-built spreadsheet version of this calculation goes wrong.
The calculator does the conversion internally, but the sensitivity is worth internalising. At 800 mm², raising D0 from 0.1 to 0.2 drops yield from about 47% to about 25% and roughly doubles cost per good die, because the number of good dies on the wafer halves.
For screening, use a range rather than one value — roughly 0.05 to 0.25 defects/cm² reflects where mature leading-edge processes tend to sit. The monolithic-versus-chiplet conclusion is far more sensitive to D0 than to wafer cost, which scales the answer linearly.
How accurate is the dies-per-wafer estimate?
The gross count uses the classic approximation πR²/S − 2πR/√(2S), with R fixed at 147 mm (a 150 mm radius wafer minus 3 mm edge exclusion) and S the die area. It is a continuous-area approximation, not a placement simulation.
It ignores scribe lanes, test structures and partial dies at the wafer edge, so real counts usually land a few percent lower. For very large dies the absolute error is small but the relative error is not, and rounding down to whole dies matters most near the bottom: when good dies come out at zero, the tool shows “—” instead of inventing a cost figure.
Treat the dies-per-wafer output as a planning estimate and reconcile it against your foundry's own wafer maps before it drives a procurement decision.
When do chiplets beat one monolithic die?
On silicon economics alone, small dies almost always win, because the defect penalty is exponential in area. Split the same 800 mm² at 0.1 defects/cm²: two 400 mm² dies yield about 68%, four 200 mm² dies about 82%, and eight 100 mm² dies about 91%. Cost per equivalent 800 mm² of good silicon falls from roughly $590 to about $370, $280 and $245 respectively.
That is silicon only, and packaging can erase it. Chiplets need a substrate or interposer, die-to-die interconnect, extra assembly steps and known-good-die test, and assembly yield compounds across the stack. If the chiplet route adds even $150 per package in interposer, assembly and test that the monolithic design would not have paid, most of the eight-chiplet advantage is gone.
Our working rule after running many comparisons: chiplets pay when the silicon saving per package is at least two to three times your best estimate of the added packaging cost. Early packaging estimates for a new flow are usually optimistic, so keep that margin rather than shaving it.
Sanity checks before you take the number to a meeting
Wafer cost is the input people under-specify most often. It should be the fully processed wafer price — processing plus mask amortisation handled the way your organisation normally handles it — not a headline number from a public price list. If masks sit in a separate budget line, say so when you present the result.
The second frequent mistake is comparing the yield percentages of a monolithic die and a chiplet set as if they were the same unit. A 91% chiplet yield and a 47% monolithic yield are not comparable, because you need eight chiplets to replace one die. Compare cost per equivalent package or cost per mm² of good silicon, both of which the tool reports.
Finally, use the output as a screening tool, not a forecast. It cannot see redundancy, repair coverage, test escapes or packaging yield, and it says nothing about whether a given chiplet partition is feasible for power and bandwidth. Verify against foundry data before the figure reaches a sourcing decision — other free utilities on OnlineFree.app follow the same one-screen, no-sign-up approach.
Frequently asked questions
How does the Die Yield And Cost Calculator compute cost per good die?
It divides the wafer cost you enter by the good dies on that wafer — predicted yield from the Murphy model multiplied by the gross dies-per-wafer estimate, then rounded down. With the defaults of 800 mm², 0.1 defects/cm² and $17,000, that gives roughly 47% yield, about 62 gross dies, about 29 good dies and near $590 per good die.
Is the Die Yield And Cost Calculator free to use?
Yes. It is a free, no-sign-up browser calculator on OnlineFree.app, with no account and no export step required. Treat it as a fast sanity check rather than a system of record, and avoid entering confidential pricing you would not put into any web form.
Why does splitting a die into chiplets improve yield so much?
Yield falls roughly exponentially with area, so one 800 mm² die at 0.1 defects/cm² yields about 47%, while eight 100 mm² chiplets of the same total area each yield about 91%. The defect saving is real, but it must cover interposer or substrate cost, assembly yield loss and extra test before the split actually saves money.
What defect density should I enter in the calculator?
Use the D0 your foundry reports for that specific process. The tool's default is 0.1 defects/cm², with roughly 0.05–0.25 described as a typical mature leading-edge range. For screening, run both 0.1 and 0.25 and compare — the chiplet-versus-monolithic answer moves much more with D0 than with wafer cost.
Does the calculator include packaging and test cost?
No. The output is bare-silicon cost: processed wafer cost divided by good dies. Bumping, substrate or interposer, assembly yield and final test all sit outside the model, and that gap matters most for chiplet comparisons, where packaging is a large share of total cost. Add those line items yourself before quoting a package price.