Chips · cost lab
64 of the largest possible dies fit on a 300 mm wafer
At the $17,500 a wafer that Epoch AI’s model uses for the NVIDIA H100/H200, NVIDIA B200 and AMD MI300X, each working die costs $607 of silicon. The defect density is assumed; type your own below.
Embed
The frame credits Epoch AI and NVIDIA and links back to its page. Preview it.
Cite
64 whole dies of 26 by 33 mm, the reticle limit
a 300 mm wafer · 3 mm edge ring · 0.1 mm scribe line- Whole dies
- 64
- The usual formula
- 56
- Wafer under a whole die
- 78%
- Good at 0.1 defects per cm²
- 29
- Silicon in each good die
- $607
Each rectangle is a whole die. The grid is placed where it fits the most, out of 576 positions tried; the usual formula, πd²/4S − πd/√2S, counts fewer. A die is good when no killing defect lands on it, so the share that works falls as the die grows and as the defect density rises.
- Whole dies a wafer
- —
- Yield
- —
- Good dies a wafer
- —
- Silicon in one good die
- —
- Logic silicon in one chip
- —
- Interposer
- —
- Memory
- —
- Logic, interposer and memory
- —
The map above shows the die you type; with chiplets, one chiplet. Assembly, test and the rest of the board are left out.
Half-size dies yield 66%, full-size dies 45%
good dies a wafer and the silicon in each, Murphy’s model, $17,500 a wafer| Die | Area | Whole | 0.05 per cm² | 0.1 per cm² | 0.2 per cm² | 0.5 per cm² |
|---|---|---|---|---|---|---|
| The reticle limit | 858 mm² | 64 | $414 42 good · 66% | $607 29 good · 45% | $1,197 15 good · 23% | $5,173 3 good · 5% |
| NVIDIA GH100 (H100)drawn square | 814 mm² | 68 | $381 46 good · 67% | $550 32 good · 47% | $1,056 17 good · 24% | $4,412 4 good · 6% |
| Half the reticledrawn square | 429 mm² | 133 | $162 108 good · 81% | $199 88 good · 66% | $292 60 good · 45% | $777 23 good · 17% |
| A 200 mm² diedrawn square | 200 mm² | 300 | $64 272 good · 91% | $71 246 good · 82% | $86 204 good · 68% | $146 120 good · 40% |
| A 100 mm² diedrawn square | 100 mm² | 612 | $30 582 good · 95% | $32 554 good · 91% | $35 503 good · 82% | $46 379 good · 62% |
NVIDIA’s 814 mm² GH100: 68 whole dies a wafer, $550 of silicon for each good one at 0.1 per cm². NVIDIA gives its area, not its sides, so it is drawn as a square.
Three ways to count what works, for a die at the reticle limit
| Defects per cm² | Poissone^(−AD) | Murphy((1 − e^(−AD)) / AD)² | Negative binomial(1 + AD/α)^(−α) |
|---|---|---|---|
| 0.05 | 65%$420 | 66%$414 | 67%$408 |
| 0.1 | 42%$645 | 45%$607 | 47%$582 |
| 0.2 | 18%$1,521 | 23%$1,197 | 26%$1,062 |
| 0.5 | 1%$19,952 | 5%$5,173 | 7%$3,924 |
A is the die’s area in cm² and D the defect density. The negative binomial lets defects cluster, with α = 3; Poisson assumes they never do.
8 chiplets cut the logic silicon a working chip needs by 50%
A reticle-size chip at 0.1 per cm², Murphy’s model, with 10% extra area on each chiplet for the links between them| Chiplets | 0% extra area | 10% extra area | 20% extra area |
|---|---|---|---|
| One die | $607 858 mm² · 64 a wafer · 45% | $607 858 mm² · 64 a wafer · 45% | $607 858 mm² · 64 a wafer · 45% |
| 2 | $398 429 mm² · 133 a wafer · 66% | $455 472 mm² · 121 a wafer · 64% | $512 515 mm² · 112 a wafer · 61% |
| 4 | $310 215 mm² · 279 a wafer · 81% | $345 236 mm² · 256 a wafer · 79% | $392 257 mm² · 230 a wafer · 78% |
| 8 | $273 107 mm² · 571 a wafer · 90% | $304 118 mm² · 518 a wafer · 89% | $334 129 mm² · 476 a wafer · 88% |
Each chiplet is tested before it is joined, so only good ones are used. Smaller dies waste less of the round wafer and fewer of them catch a defect; the links between chiplets cost area, and assembly can lose some. Type your own split above.
An interposer 3.3 times the reticle field costs $750 at $12,000 a wafer
the interposer under the dies and the memory, by size| Size | Area | A wafer | At $12,000 a wafer | At $17,000 a wafer |
|---|---|---|---|---|
| 1 × the reticle field | 858 mm² | 62 | $194 | $274 |
| 2 × the reticle field | 1,716 mm² | 29 | $414 | $586 |
| 3.3 × the reticle field | 2,831 mm² | 16 | $750 | $1,063 |
| 5 × the reticle field | 4,290 mm² | 12 | $1,000 | $1,417 |
An interposer larger than one reticle field is printed in pieces that join. Each is drawn as a square and counted as good. Epoch AI’s model uses $12,000 a wafer for Amazon Trainium1, Amazon Trainium2, AMD MI300A and 11 more, and $17,000 for NVIDIA B200 and NVIDIA B300. The H100 carries 5 stacks of HBM; no free source publishes a price per stack, so the calculator takes yours.
The wafer prices Epoch AI’s model uses
| Chip | Logic wafer | Interposer wafer | Quarters |
|---|---|---|---|
| Amazon Trainium1 | — | $12,000 | 4 |
| Amazon Trainium2 | $17,500 | $12,000 | 9 |
| AMD MI300A | $17,500 | $12,000 | 5 |
| AMD MI300X | $17,500 | $12,000 | 9 |
| AMD MI308X | $17,500 | $12,000 | 4 |
| AMD MI325X | $17,500 | $12,000 | 6 |
| AMD MI350X | $20,000 | $12,000 | 5 |
| AMD MI355X | $20,000 | $12,000 | 4 |
| Google TPU v5e | $17,500 | $12,000 | 5 |
| Google TPU v5p | $17,500 | $12,000 | 6 |
| Google TPU v6e | $17,500 | $12,000 | 8 |
| Google TPU v7 | $20,000 | $12,000 | 4 |
| NVIDIA B200 | $17,500 | $17,000 | 6 |
| NVIDIA B300 | $17,500 | $17,000 | 4 |
| NVIDIA H100/H200 | $17,500 | $12,000 | 8 |
| NVIDIA H20 | $17,500 | $12,000 | 7 |
Wafer prices are Epoch AI’s modelled estimates, not prices a foundry has published. Defect density, edge and scribe are assumptions, each shown where it is used. Epoch’s medians, Q1 2024 to Q1 2026.
Sources
cost.csv Every dataset → Status → Glossary →- Epoch AI, Data on AI chip components, by chip, retrieved 19 Oct 2026. CC BY 4.0. A wafer price is Epoch’s cost over Epoch’s wafers for each chip.
- NVIDIA, NVIDIA Hopper Architecture In-Depth, 22 Mar 2022: the GH100’s area and memory stacks.
- Carl Zeiss SMT, Optics for EUV Lithography (2018 EUVL Workshop), 13 Jun 2018: the 26 by 33 mm field.
- Assumed: the edge ring, the scribe line, the defect densities, α and the extra area per chiplet. Each is a field above.