Silicon Analysts
Loading...

How Many Chips Fit on a Wafer?

The number of chips per wafer is the single biggest driver of semiconductor manufacturing cost. A small IoT chip yields thousands of dies per 300mm wafer, while a massive AI accelerator die like the NVIDIA H100 yields fewer than 80. Here’s how to calculate it and why it matters.

Quick Reference: Dies Per Wafer by Chip Type

Chip TypeDie Size (mm²)Process NodeGDPW (300mm)Est. YieldNet Good Dies
Small IoT MCU25 (5×5)28nm~2,800~95%~2,660
Smartphone SoC120 (10×12)4nm~580~85%~493
Laptop CPU200 (13×15)5nm~340~80%~272
NVIDIA H100 GPU814 (26×31)4nm~80~65%~52
Apple M4 Ultra~400 (20×20)3nm~160~72%~115
AMD MI300X (chiplet)~2005nm~340~80%~272

GDPW values assume 300mm wafer with 3mm edge exclusion. Yields are estimates based on published defect density data for each node’s maturity level.

The GDPW Formula

GDPW = (π × (D/2 − E)²) / S − (π × (D/2 − E)) / √(2 × S)

Where:

  • D = wafer diameter (300mm for modern fabs)
  • E = edge exclusion zone (typically 2–3mm)
  • S = die area (width × height in mm²)

The first term estimates how many rectangles fit in the usable circular area. The second term subtracts partial dies lost along the wafer edge. This is an approximation — actual die placement uses stepping algorithms that can recover a few extra dies depending on die aspect ratio and orientation.

Why Die Size Matters So Much for Cost

Die size has a non-linear impact on chip cost. Doubling die area does not simply double cost — it more than doubles it, for two compounding reasons:

  1. Fewer dies per wafer: On a circular wafer, doubling the rectangular die area yields less than half the number of dies due to edge losses.
  2. Lower yield: Larger dies are more likely to contain a killer defect. Yield drops exponentially with die area at any given defect density.

This is precisely why the semiconductor industry is moving toward chiplet architectures — multiple small dies connected via advanced packaging (CoWoS, EMIB, UCIe) instead of one monolithic die. AMD’s MI300X uses this approach: multiple ~200mm² chiplets instead of a single 800mm²+ die, dramatically improving yield and cost.

Yield: Why You Don’t Get All the Dies

GDPW tells you how many dies physically fit on a wafer. But not all of them work. Defects introduced during fabrication — particles, lithography errors, implant variations — kill some fraction of dies. The percentage that pass testing is called die yield.

Yield is typically modeled using Murphy’s model or the Poisson model. The key inputs are:

  • Defect density (D0): Typically 0.05–0.15 defects/cm² for mature nodes, 0.2–0.5 for leading edge
  • Die area (A): Larger dies catch more defects

At TSMC’s 3nm node, a small 50mm² die might achieve 90%+ yield, while an 800mm² GPU die could see yields below 50% in early production. This yield gap is the main reason large AI accelerator dies are so expensive per unit.

How AI Chips Changed the Economics

The NVIDIA H100’s 814mm² die pushed the limits of what’s economically viable on a single monolithic chip. At ~80 gross dies per wafer and an estimated 60–70% yield, each wafer produces only ~50 good H100 dies. With TSMC N4 wafers costing ~$16,000+, the die cost alone exceeds $300 — before packaging, HBM memory, and test.

NVIDIA’s Blackwell B200 moved to a dual-die architecture to address this: two smaller dies connected via NVLink, improving yield while delivering more compute. This chiplet trend is now standard across the industry for high-performance AI accelerators.

Try It Yourself

Enter your die dimensions and process node to calculate GDPW, yield, and full chip cost with our free interactive calculator.

Open Chip Cost Calculator →

Related Resources

Chips Per Wafer FAQ

How many chips fit on a 300mm wafer?
It depends on die size. A small IoT chip (5×5mm, 25mm²) yields ~2,800 gross dies. A smartphone SoC (10×12mm, 120mm²) yields ~580. A laptop CPU (13×15mm, 200mm²) yields ~340. A large GPU like the NVIDIA H100 (26×31mm, 814mm²) yields only ~80 gross dies. Net good dies after yield losses are typically 60–95% of gross.
What is GDPW (Gross Dies Per Wafer)?
GDPW is the theoretical maximum number of rectangular dies that can be cut from a circular silicon wafer. The formula accounts for wafer diameter (typically 300mm), die area (width × height), and edge exclusion zones (2–3mm). GDPW is the starting point for all chip cost calculations.
How is die yield calculated?
Die yield is calculated using defect density models such as Murphy’s, Poisson, or negative binomial. The key inputs are defect density (defects/cm²) and die area (cm²). Larger dies have lower yield because they are more likely to contain a killer defect. Typical yields range from 95% for small mature-node chips to below 50% for very large leading-edge dies.
Why does doubling die size more than double chip cost?
Doubling die area has two compounding effects: (1) you get fewer than half the gross dies per wafer due to the circular wafer geometry, and (2) yield drops non-linearly because larger dies are more likely to contain defects. A 400mm² die can cost 3–4x as much per good die as a 200mm² die on the same process.
What is the GDPW formula?
The standard GDPW formula is: GDPW = (π × (D/2 − E)²) / S − (π × (D/2 − E)) / √(2 × S), where D = wafer diameter (300mm), E = edge exclusion (2–3mm), and S = die area (width × height in mm²). This approximation accounts for the inefficiency of packing rectangles on a circle.