chip cost

**Semiconductor Economics: Chip, Wafer, and Fab Costs** **Overview** Semiconductor economics operates across three interconnected cost levels, each driving the next in a hierarchical structure that determines the final price of every chip. --- **1. Fab (Fabrication Plant) Cost** The foundation of semiconductor economics—the capital expenditure required to build and equip a fabrication facility. **Capital Expenditure Breakdown** - **Modern leading-edge fabs (3nm/2nm):** $15–25+ billion to construct - **Historical comparison:** - Year 2000: ~$1–2 billion per fab - Year 2010: ~$3–5 billion per fab - Year 2020: ~$10–15 billion per fab - Year 2024+: ~$20–30 billion per fab **Cost Components** - **Equipment (70–80% of capital cost):** - ASML EUV lithography machines: ~$350–400 million each - Deposition tools (CVD, PVD): $5–20 million each - Etching systems: $5–15 million each - Metrology and inspection: $2–10 million each - Ion implantation: $3–8 million each - **Facility construction (20–30% of capital cost):** - Cleanroom (Class 1-10): $3,000–5,000 per square foot - Ultra-pure water systems: $100–500 million - Vibration isolation foundations - Chemical delivery systems - HVAC and air filtration **Depreciation Model** Fab equipment is typically depreciated over 5–7 years: $$ \text{Annual Depreciation} = \frac{\text{Fab Capital Cost}}{\text{Depreciation Period}} $$ **Example:** $$ \text{Annual Depreciation} = \frac{\$20 \text{ billion}}{5 \text{ years}} = \$4 \text{ billion/year} $$ --- **2. Wafer Cost** The cost to process a single silicon wafer (typically 300mm diameter) through hundreds of manufacturing steps. **Wafer Cost by Process Node** | Node | Approximate Wafer Cost | Typical Applications | |------|------------------------|---------------------| | 3nm | $18,000–$22,000 | Flagship mobile SoCs, high-end GPUs | | 5nm | $16,000–$18,000 | Premium smartphones, AI accelerators | | 7nm | $10,000–$12,000 | Gaming consoles, data center CPUs | | 14nm | $5,000–$7,000 | Mid-range processors, FPGAs | | 28nm | $3,000–$4,000 | Automotive, WiFi, Bluetooth | | 65nm | $2,000–$2,500 | MCUs, power management | | 180nm | $1,000–$1,500 | Analog, sensors, legacy | **Wafer Cost Formula** $$ C_{\text{wafer}} = C_{\text{depreciation}} + C_{\text{materials}} + C_{\text{labor}} + C_{\text{utilities}} + C_{\text{overhead}} $$ Where: - $C_{\text{depreciation}}$ = Equipment depreciation per wafer - $C_{\text{materials}}$ = Silicon, photoresists, gases, chemicals, CMP slurries - $C_{\text{labor}}$ = Engineering and technician costs - $C_{\text{utilities}}$ = Electricity, ultra-pure water, gases - $C_{\text{overhead}}$ = Maintenance, yield engineering, facility costs **Wafer Throughput Economics** $$ C_{\text{depreciation/wafer}} = \frac{\text{Annual Depreciation}}{\text{Wafers per Year}} $$ **Example for a $20B fab producing 100,000 wafers/month:** $$ C_{\text{depreciation/wafer}} = \frac{\$4 \text{ billion/year}}{1.2 \text{ million wafers/year}} \approx \$3,333 \text{ per wafer} $$ --- **3. Chip (Die) Cost** The cost per individual chip, derived from wafer economics and manufacturing yield. **Fundamental Die Cost Equation** $$ C_{\text{die}} = \frac{C_{\text{wafer}}}{N_{\text{dies}} \times Y} $$ Where: - $C_{\text{die}}$ = Cost per good die - $C_{\text{wafer}}$ = Total wafer processing cost - $N_{\text{dies}}$ = Number of dies per wafer (gross) - $Y$ = Yield (fraction of functional dies) **Dies Per Wafer Calculation** For a circular wafer with rectangular dies: $$ N_{\text{dies}} \approx \frac{\pi \times D^2}{4 \times A_{\text{die}}} - \frac{\pi \times D}{\sqrt{2 \times A_{\text{die}}}} $$ Where: - $D$ = Wafer diameter (300mm for modern fabs) - $A_{\text{die}}$ = Die area in mm² **Simplified approximation:** $$ N_{\text{dies}} \approx \frac{\pi \times (150)^2}{A_{\text{die}}} \times 0.85 $$ The 0.85 factor accounts for edge losses and scribe lines. **Dies Per Wafer Examples** | Die Size (mm²) | Approximate Dies/Wafer | Example Chips | |----------------|------------------------|---------------| | 5 | ~12,000 | Small MCUs, sensors | | 25 | ~2,400 | Bluetooth, WiFi chips | | 100 | ~600 | Mobile SoCs, mid-range GPUs | | 300 | ~200 | Desktop CPUs, gaming GPUs | | 600 | ~90 | Data center GPUs | | 800 | ~60 | Large AI accelerators (H100) | | 1,200 | ~35 | Largest monolithic dies | **Yield Models** **Murphy's Yield Model** $$ Y = \left( \frac{1 - e^{-D_0 \times A}}{D_0 \times A} \right)^2 $$ **Poisson Yield Model (simpler)** $$ Y = e^{-D_0 \times A} $$ Where: - $Y$ = Die yield (fraction) - $D_0$ = Defect density (defects per cm²) - $A$ = Die area (cm²) **Typical defect densities:** - Mature process: $D_0 \approx 0.05–0.1$ defects/cm² - New process (early): $D_0 \approx 0.3–0.5$ defects/cm² - New process (ramping): $D_0 \approx 0.1–0.2$ defects/cm² **Yield Impact Examples** For a 600mm² die ($A = 6$ cm²): **Mature process** ($D_0 = 0.1$): $$ Y = e^{-0.1 \times 6} = e^{-0.6} \approx 0.55 = 55\% $$ **Early production** ($D_0 = 0.3$): $$ Y = e^{-0.3 \times 6} = e^{-1.8} \approx 0.17 = 17\% $$ --- **4. Complete Cost Model** **Total Manufacturing Cost Per Chip** $$ C_{\text{total}} = C_{\text{die}} + C_{\text{packaging}} + C_{\text{testing}} + C_{\text{design\_amort}} $$ Where: $$ C_{\text{design\_amort}} = \frac{C_{\text{NRE}}}{\text{Total Units Produced}} $$ - $C_{\text{NRE}}$ = Non-Recurring Engineering costs (design, masks, validation) **NRE Costs by Node** | Node | Approximate NRE Cost | |------|---------------------| | 3nm | $500M – $1B+ | | 5nm | $400M – $700M | | 7nm | $250M – $400M | | 14nm | $100M – $200M | | 28nm | $50M – $100M | | 65nm | $20M – $40M | **Packaging Costs** - **Standard wire bond:** $0.10 – $1.00 - **Flip chip BGA:** $2 – $10 - **Advanced fan-out (InFO):** $10 – $50 - **2.5D interposer (CoWoS):** $100 – $400 - **3D stacking:** $200 – $600+ --- **5. Worked Examples** **Example 1: AI Accelerator Chip** **Parameters:** - Node: TSMC 5nm - Die size: 600mm² - Wafer cost: $17,000 - Defect density: $D_0 = 0.12$ /cm² **Calculations:** **Dies per wafer:** $$ N_{\text{dies}} = \frac{\pi \times 150^2}{600} \times 0.85 \approx 100 \text{ dies} $$ **Yield:** $$ Y = e^{-0.12 \times 6} \approx e^{-0.72} \approx 0.49 = 49\% $$ **Die cost:** $$ C_{\text{die}} = \frac{\$17,000}{100 \times 0.49} = \frac{\$17,000}{49} \approx \$347 $$ **Total chip cost:** $$ C_{\text{total}} = \$347 + \$250_{\text{(CoWoS)}} + \$30_{\text{(test)}} + \$50_{\text{(design)}} \approx \$677 $$ --- **Example 2: IoT Microcontroller** **Parameters:** - Node: 40nm - Die size: 5mm² - Wafer cost: $3,000 - Defect density: $D_0 = 0.05$ /cm² **Calculations:** **Dies per wafer:** $$ N_{\text{dies}} = \frac{\pi \times 150^2}{5} \times 0.85 \approx 12,000 \text{ dies} $$ **Yield:** $$ Y = e^{-0.05 \times 0.05} \approx e^{-0.0025} \approx 0.997 = 99.7\% $$ **Die cost:** $$ C_{\text{die}} = \frac{\$3,000}{12,000 \times 0.997} \approx \$0.25 $$ **Total chip cost:** $$ C_{\text{total}} = \$0.25 + \$0.15_{\text{(pkg)}} + \$0.05_{\text{(test)}} + \$0.05_{\text{(design)}} \approx \$0.50 $$ --- **6. Economic Dynamics** **Learning Curve Effect** Manufacturing cost decreases with cumulative volume: $$ C_n = C_1 \times n^{-b} $$ Where: - $C_n$ = Cost at cumulative unit $n$ - $C_1$ = Cost of first unit - $b$ = Learning exponent (typically 0.1–0.3 for semiconductors) - Learning rate = $2^{-b}$ (typically 85–95%) **Economies of Scale** **Fab utilization impact:** $$ C_{\text{wafer}}(\text{util}) = \frac{C_{\text{fixed}}}{\text{util}} + C_{\text{variable}} $$ - At 50% utilization: costs ~1.5× baseline - At 90% utilization: costs ~1.05× baseline - At 100% utilization: minimum cost achieved **Cost Sensitivity Analysis** **Die cost sensitivity to yield:** $$ \frac{\partial C_{\text{die}}}{\partial Y} = -\frac{C_{\text{wafer}}}{N_{\text{dies}} \times Y^2} $$ For large, expensive dies, yield improvements have dramatic cost impacts. --- **7. Industry Structure Implications** **Why Only 3 Companies at Leading Edge** **Minimum efficient scale calculation:** $$ \text{Revenue Required} = \frac{\text{Annual CapEx} + \text{R\&D}}{\text{Margin}} $$ $$ \text{Revenue Required} \approx \frac{\$15B + \$5B}{0.40} = \$50B+ \text{ annually} $$ Only TSMC, Samsung, and Intel can sustain this investment level. **Foundry Model Economics** **Fabless company advantage:** $$ \text{ROI}_{\text{fabless}} = \frac{\text{Chip Revenue} - \text{Foundry Cost} - \text{Design Cost}}{\text{Design Cost}} $$ **IDM (Integrated Device Manufacturer):** $$ \text{ROI}_{\text{IDM}} = \frac{\text{Chip Revenue} - \text{Mfg Cost} - \text{Design Cost}}{\text{Fab CapEx} + \text{Design Cost}} $$ The fabless model eliminates fab capital from the denominator, enabling higher ROI for design-focused companies. --- **8. Summary Equations** **Core Formulas Reference** | Metric | Formula | |--------|---------| | Die Cost | $C_{\text{die}} = \frac{C_{\text{wafer}}}{N_{\text{dies}} \times Y}$ | | Dies per Wafer | $N \approx \frac{\pi r^2}{A_{\text{die}}} \times 0.85$ | | Poisson Yield | $Y = e^{-D_0 \times A}$ | | Total Cost | $C_{\text{total}} = C_{\text{die}} + C_{\text{pkg}} + C_{\text{test}} + C_{\text{NRE}}$ | | Depreciation/Wafer | $C_{\text{dep}} = \frac{\text{CapEx}/t}{\text{WPY}}$ | | Learning Curve | $C_n = C_1 \times n^{-b}$ | --- **9. Current Market Dynamics (2024–2025)** **Key Trends** - **AI demand:** Consuming 20%+ of advanced node capacity - **Geopolitical reshoring:** Adding 20–30% cost premium for non-Taiwan fabs - **EUV bottleneck:** ASML's monopoly constrains expansion - **Advanced packaging:** Becoming equal cost driver to node shrinks - **Chiplet economics:** Enabling yield improvement through smaller dies **Government Subsidies Impact** - **US CHIPS Act:** $52B in subsidies - **EU Chips Act:** €43B in public/private investment - **Effect:** Artificially reducing effective CapEx for new fabs --- *Document generated: January 2025* *Data sources: Industry reports, foundry pricing estimates, public financial disclosures*

Go deeper with CFSGPT

Get AI-powered deep-dives, save terms, and run advanced simulations — free account.

Create Free Account