From Circuit to Chip

From Circuit to Chip

Think about a rubber stamp. You carve a design into it, press it onto an ink pad, then press it onto paper — and there's your image. Press it a thousand times, and you have a thousand identical copies. The cost per copy is essentially zero beyond the initial carving.

That's the fundamental insight behind chip manufacturing. Instead of stamping ink onto paper, a chip factory stamps circuit patterns onto silicon using light. Instead of millimeter precision, it works at nanometer precision — finer than a human hair by a factor of ten thousand. And instead of a thousand copies, it makes millions of identical chips from a single design.

Core Principles

From Discrete Components to Integrated Circuits

Early computers were assemblies of individual discrete components — resistors, capacitors, and transistors soldered onto circuit boards and connected by wires. If you wanted to double the number of transistors, you had to physically solder twice as many parts. Machines grew larger, heavier, and harder to maintain. ENIAC weighed 27 tons and filled a room.

In 1958, Jack Kilby at Texas Instruments had a breakthrough idea: fabricate multiple electronic components on the same piece of semiconductor material, forming all connections in a single manufacturing process. This was the birth of the integrated circuit (IC). Almost simultaneously, Robert Noyce at Fairchild Semiconductor independently arrived at the same concept and solved the interconnection problem using a planar process. Kilby received the 2000 Nobel Prize in Physics for this work.

The revolution wasn't just miniaturization — it was that circuits were no longer assembled, they were printed. The cost per transistor dropped with every new process generation, and the pace of improvement compounded exponentially.

Photolithography: Writing Circuits with Light

The heart of modern chip manufacturing is photolithography — literally "writing on stone with light." Here's the essential process:

Step 1: Coat a silicon wafer with photoresist
        (a chemical that changes when exposed to light)

        ┌─────────────────────────────┐
        │  Photoresist (thin layer)   │
        ├─────────────────────────────┤
        │  Silicon wafer              │
        └─────────────────────────────┘

Step 2: Shine extreme ultraviolet (EUV) light through a mask
        (the mask is the "stamp" — opaque in some places, transparent in others)

        [EUV Light Source]
               │
           [Photomask] ← circuit pattern etched into glass
               │ light passes through transparent regions
               ↓
        ┌─────────────────────────────┐
        │ Exposed regions chemically  │
        │ altered by light            │
        ├─────────────────────────────┤
        │  Silicon wafer              │
        └─────────────────────────────┘

Step 3: Develop — wash away exposed (or unexposed) photoresist
Step 4: Etch — remove unprotected silicon with chemicals or plasma
Step 5: Dope — implant atoms to create N-type or P-type semiconductor
Step 6: Deposit — lay down metal (copper or aluminum) for wiring
Step 7: Repeat steps 1–6 dozens of times to build a multilayer circuit

A modern chip has 10 to 15 metal layers stacked on top of each other, each requiring a separate lithography pass. The entire fabrication process takes two to three months and involves hundreds of distinct steps. The silicon wafer that enters the fab is worth a few hundred dollars; the chips that come out can be worth millions.

Moore's Law: The Doubling That Defined an Era

In 1965, Gordon Moore — co-founder of Intel — noticed an empirical pattern: the number of transistors on an integrated circuit roughly doubled every two years (later revised to 18 months), while the cost per transistor continued to fall. This observation became known as Moore's Law.

Transistor count (log scale)

10^11 |                                        ● Modern CPU (100B+)
      |                                  ●
10^10 |
      |                        ●
10^9  |              ●
      |     ●
10^8  |●
      |
10^7  |
      +──────────────────────────────────────→ Year
       1970   1980   1990   2000   2010   2020

Every ~18 months: count doubles → straight line on log scale

Moore's Law was never a physical law — it was an engineering and business commitment. Chip companies invested to maintain the pace because falling behind meant losing market share. The prophecy was self-fulfilling for over 50 years, driving the most sustained technological improvement in human history.

Process Nodes: What Do the Nanometer Numbers Mean?

You've seen headlines about 7nm, 5nm, 3nm chips. The number originally referred to the gate length of a transistor — a key physical dimension. Today it's more of a marketing label, with different manufacturers using different measurement conventions. But the direction is clear: smaller number means smaller transistors, more of them per square millimeter, and lower power consumption.

Historical process node evolution:

1971   Intel 4004        10,000 nm       2,300 transistors
1989   Intel 486          1,000 nm   1,200,000 transistors
2003   Pentium 4            130 nm  55,000,000 transistors
2012   Ivy Bridge (1st FinFET)  22 nm   1,400,000,000
2019   Apple A13              7 nm   8,500,000,000
2022   Apple A16              4 nm  16,000,000,000
2023   Apple A17 Pro          3 nm  19,000,000,000

At 3nm, one transistor is approximately:
- 15–20 silicon atoms wide
- ~30× smaller than the COVID-19 virus
- ~100× smaller than visible light wavelengths
  (which is why ordinary light can't be used to print it)

Why Can We Keep Shrinking?

Diffraction is the fundamental obstacle: you can't use light to print features smaller than the light's own wavelength. Visible light runs from 400–700nm — far too coarse for 7nm circuits. The industry addressed this on multiple fronts:

Switch to shorter wavelengths: Extreme Ultraviolet (EUV) light has a wavelength of just 13.5nm. Generating it requires firing a laser at tin droplets 50,000 times per second, creating a plasma that emits EUV radiation. The machine that does this — built exclusively by Dutch company ASML — costs over $200 million per unit.

Multi-patterning: Split a single circuit layer into multiple separate exposures that are precisely overlaid. The combined result is finer than any single exposure could achieve. The downside is that each extra pass multiplies cost and time.

3D stacking: Rather than cramming more into a flat plane, Intel (Foveros), TSMC (SoIC), and Samsung (X-Cube) now stack chips vertically, connecting them with dense arrays of microscopic pillars called through-silicon vias (TSVs). Building upward like a skyscraper instead of outward like a suburb.

Hands-On Verification

Explore Moore's Law as exponential growth in Python:

import math

def moores_law(start_count, years, doubling_months=18):
    """Predict transistor count based on Moore's Law."""
    doublings = years * 12 / doubling_months
    return start_count * (2 ** doublings)

# Historical data points
historical = {
    1971: 2_300,           # Intel 4004
    1989: 1_200_000,       # Intel 486
    2003: 55_000_000,      # Pentium 4
    2012: 1_400_000_000,   # Ivy Bridge (1st FinFET)
    2019: 8_500_000_000,   # Apple A13
    2022: 16_000_000_000,  # Apple A16
}

start_year = 1971
start_count = 2_300

print(f"{'Year':6} {'Predicted (Moore)':>20} {'Actual':>20} {'Ratio':>8}")
print("-" * 58)
for year, actual in sorted(historical.items()):
    predicted = moores_law(start_count, year - start_year)
    ratio = actual / predicted
    print(f"{year:6} {predicted:>20,.0f} {actual:>20,} {ratio:>8.2f}x")

# How long to go from 2,300 to 16,000,000,000 transistors?
doublings_needed = math.log2(16_000_000_000 / 2_300)
months_needed = doublings_needed * 18
print(f"\nDoublings from 4004 to A16: {doublings_needed:.1f}")
print(f"At 18 months each: {months_needed:.0f} months = {months_needed/12:.0f} years")
print(f"Actual span: {2022 - 1971} years — Moore's Law slightly overestimates the pace")

Output (approximate):

Year     Predicted (Moore)               Actual    Ratio
----------------------------------------------------------
1971                 2,300                2,300     1.00x
1989            10,240,000            1,200,000     0.12x
2003           939,524,096           55,000,000     0.06x
2012        17,179,869,184        1,400,000,000     0.08x
...

Doublings from 4004 to A16: 22.8
At 18 months each: 410 months = 34 years
Actual span: 51 years — Moore's Law slightly overestimates the pace

The prediction diverges from reality — Moore's Law was never precise. But the exponential direction held for half a century, which is remarkable for any empirical engineering trend.

🔬 Going Deeper

ASML and the EUV Monopoly

The EUV lithography machine is arguably the most complex manufactured object in existence. ASML, headquartered in Eindhoven, Netherlands, is the sole manufacturer, and its most advanced machine — the High-NA EUV — took decades and billions of dollars to develop. Generating EUV light requires a CO₂ laser firing at tin droplets 50,000 times per second; the resulting plasma emits 13.5nm radiation that is then focused and reflected (not transmitted, since no material is transparent at EUV wavelengths) through a system of ultra-smooth mirrors. The entire machine has about 100,000 parts, weighs 180 tons, and ships in dozens of cargo containers. It represents the cumulative engineering of hundreds of suppliers across dozens of countries — which is why restricting its export became a central lever in geopolitical semiconductor competition.

Is Moore's Law Dead?

Classical Moore's Law — transistor density doubling every 18 months at constant cost — has slowed measurably since the mid-2010s. But "death" is too simple. The industry is transitioning from a single dimension of improvement (shrinking transistors) to multiple parallel tracks: 3D integration (Intel Foveros, TSMC SoIC), new transistor architectures (Gate-All-Around, or GAA, introduced at Samsung's 3nm node), new channel materials (2D materials like molybdenum disulfide may extend scaling below 2nm), and disaggregation (chiplets — assembling CPUs from specialized dies rather than a single monolithic chip). Performance improvements continue; they've simply become more expensive and more multidimensional.

Recommended Reading

• Chip War (Chris Miller, 2022) — A comprehensive history from the transistor's invention through the current semiconductor geopolitical struggle. It reads like a thriller and provides essential context for understanding why chipmaking is now at the center of international relations.
• The Chip (T.R. Reid) — Focuses on the human story of Kilby and Noyce's parallel invention of the integrated circuit. If you want to understand the culture and personalities that built Silicon Valley, this is the starting point.