A 22% Smaller Die Does Not Mean 22% Lower Cost
A 22% smaller die is worth roughly 25% off cost per good die, but only if the wafer price stays put. Push the wafer price up more than 32.6% and the entire area gain disappears. That threshold is the whole argument, and almost no announcement gives you the numbers to locate it.
KEY TAKEAWAYS
1. Cutting die area 22% lifts good die per 300mm wafer by 32.6%, not 22%. Edge loss and defect yield both improve alongside the shrink.
2. The break-even is a 32.6% wafer price increase. Above that, cost per good die rises despite the smaller die.
3. Samsung's own published figure for the SF3 to SF2 node step is 5% area. The 22% NPU number circulating this month comes from a trade report with no stated baseline.
Why a smaller die pays more than proportionally
A 300mm wafer gives you about 70,686 mm² of usable silicon. At 100 mm² per die the arithmetic says 706 dies. Real placement gets about 640, because dies straddling the round edge are scrap.
Drop the die to 78 mm² and you get about 831. The area came down 22%. The die count went up 29.8%. Pure area accounts for 28.2% of that (1 divided by 0.78) and the rest is edge waste you no longer pay for.
Defect yield adds to it. A smaller die is a smaller target for a killer defect. Under the Murphy yield model at an assumed defect density of 0.1 per cm², yield moves from 90.6% to 92.5%. Good die per wafer goes from 580 to 769, up 32.6%.
What is actually published, and what is not
The 22% figure comes from a DIGITIMES report dated September 10, 2026, describing an NPU implemented with Synopsys Fusion Compiler on Samsung Foundry's SF2P process. I could not find that number in either company's own materials.
The closest primary source is Synopsys's June 16, 2025 release on certified design flows for SF2P, SF2 and SF4X. The quantified claims there are a 10x turnaround reduction using 3DIC Compiler, a 6% improvement in worst-case eye opening and HBM routing completed in four hours. No area figure.
Samsung's own published 2nm PPA, from Samsung Foundry Forum 2023, is 12% higher performance, 25% better power efficiency and 5% smaller area for SF2 against SF3. That is the honest reference point for what a node step buys you.
The wafer price step that erases the gain
Cost per good die is wafer price divided by good die count. A 32.6% increase in good die is a 24.6% reduction in cost per die, holding wafer price constant.
Wafer price does not hold constant across a node transition. It goes up. So the break-even is exactly the good-die gain: if the new node charges more than 32.6% per wafer, the 22% area cut buys nothing and you are paying more per shipped chip than before.
This is where the SF2P case is interesting rather than routine. The reported 22% is not a node move. It is a design-side reduction inside one process. If that holds, the wafer price is fixed and the saving stays in the cost line instead of being handed back to the foundry.
Three conditions before area becomes cost
The baseline. 22% against what? The same RTL through a different flow, the previous NPU generation or a different node are three completely different claims. The report does not say.
The verification stage. Synthesis and place-and-route results are not production silicon. Design complete, tape-out, sampling, qualified and volume are distinct stages and the difference between them is where most optimistic PPA numbers go to die. Which stage this figure describes is not disclosed.
The design ecosystem. Foundry customers do not choose a node on process specification alone. They need certified IP and a certified flow to tape out on schedule. That is what Synopsys chose to quantify in its own 2025 release, and it is a better read on competitive position than any single area number.
None of this makes area reduction into design wins. Area is one input to cost. Wins are decided by capacity, schedule, price and accumulated trust.
What I actually watch
| Signal | Why it matters |
|---|---|
| The 22% reappearing in a Synopsys or Samsung release with a stated baseline | Converts a trade report into a citable figure |
| A disclosed customer tape-out on SF2P | Moves the story from design study to committed silicon |
| Samsung's quarterly foundry segment result | The only public read on whether leading-edge mix is actually shifting |
| Which nodes the EDA vendors certify flows for | Certification tends to lead customer commitments |
Value chain read-through
| Segment | Read-through |
|---|---|
| EDA and design IP | Area and power optimization sells against process specification. Certified flows gate how fast a foundry converts interest into tape-outs |
| Foundry | Shrinking a customer's design inside an existing node lowers their cost without giving up wafer price |
| Fabless with mature design assets | Room to improve cost without a node migration, offset by redesign expense and schedule |
| Back end and packaging | A smaller die changes package and thermal design constraints alongside it |
Risks to this view
• Every die count, yield and cost figure in this post is an illustrative calculation. Actual defect density and wafer pricing are not published.
• The 22% is a secondary report. I could not verify the original announcement, and a disclosed baseline could change the reading entirely.
• Murphy is a simplified yield model. Real yield also carries lithography, packaging and burn-in losses.
• Area improvement does not imply design wins or share gains. This post does not argue that it does.
A 22% area number tells you nothing about cost on its own. It becomes readable only once you know the baseline, whether the reduction came from a node move or from design inside one node and which verification stage produced it.
Sources: DIGITIMES, September 10, 2026 (https://www.digitimes.com/news/a20260910VL218/synopsys-samsung-npu-2nm-design.html); Synopsys news release, June 16, 2025 (https://news.synopsys.com/2025-06-16-Synopsys-Accelerates-AI-and-Multi-Die-Design-Innovation-on-Advanced-Samsung-Foundry-Processes); Samsung Foundry Forum 2023 (https://semiconductor.samsung.com/news-events/news/samsung-electronics-unveils-foundry-vision-in-the-ai-era-at-samsung-foundry-forum-2023/); Samsung Foundry Forum 2024 (https://semiconductor.samsung.com/news-events/news/samsung-showcases-ai-era-vision-and-latest-foundry-technologies-at-sff-2024/). Everything here is from public sources.
Comments
Post a Comment