Visual walkthrough — Copper damascene process

We will never use a symbol before drawing it. Let's go.

---

Step 1 — What even is a wire? (picture a ditch full of copper)

WHAT. In the damascene process a wire is literally a rectangular bar of copper sitting inside a ditch carved in glass. Three measurements fully describe that bar:

• its length $L$ — how far the signal has to travel end to end,
• its width $w$ — how wide the ditch is,
• its thickness $t$ — how deep the copper fills.

WHY these three. Everything about how fast a signal crosses the wire comes from these three numbers plus the material. Nothing else. So we name them once, anchor them to the picture, and reuse them forever.

PICTURE. Below, the copper bar is drawn as a box. The green arrow is the direction current flows (along $L$). The shaded end-face is the cross-section the current has to squeeze through.

---

Step 2 — Building resistance from crowding

WHAT. Resistance $R$ measures how hard it is to push electric current through the bar. Think of a corridor full of people trying to walk from one end to the other.

WHY this picture. Two things obviously make the corridor harder to cross:

1. A longer corridor (bigger $L$) → more distance to fight through → more resistance.
2. A narrower doorway (smaller $A$) → everyone crowds → more resistance.

So resistance must grow with $L$ and shrink with $A$. The simplest formula obeying both is $R\propto L/A$.

PICTURE. Two corridors side by side: a long/thin one (crowded, high $R$) and a short/wide one (roomy, low $R$).

---

Step 3 — A worked resistance, so the numbers feel real

WHAT. Let's plug real chip dimensions into $R=\rho L/(wt)$ for a copper line.

WHY. Symbols lie until you touch a number. Doing one careful calculation also teaches the single most dangerous trap: unit conversion ($\mu \text{m}\to \text{m}$).

PICTURE. The same copper bar, now labelled with actual sizes, and the conversion staircase from micrometres to metres.

Take $L=1000\mu \text{m}$, $w=0.1\mu \text{m}$, $t=0.2\mu \text{m}$. One micrometre is $10^{-6}$ metre, so:

$L=10^{-3}\text{m},A=wt=(0.1\times 10^{-6})\times (0.2\times 10^{-6})=2\times 10^{-14}{\text{m}}^2$

• $\times 10^{-12}$ ::: because two lengths ($w$ and $t$) each carry a $10^{-6}$, and $10^{-6}\times 10^{-6}=10^{-12}$

Now plug in:

$R=\rho \frac{L}{A}=\frac{1.68\times 10^{-8}\times 10^{-3}}{2\times 10^{-14}}=840\Omega$

Read it: a single micron-scale copper wire is already $840\Omega$ — real resistance you must fight. Aluminum with the same shape would be $\frac{2.65}{1.68}\times 840\approx 1325\Omega$. Already you feel copper winning.

---

Step 4 — Where capacitance sneaks in (three conductors, not two)

WHAT. A wire never sits alone. A neighbour wire runs beside it, separated by a gap $s$ of insulator — but there is also a third conductor: the grounded silicon substrate underneath, a distance $h$ below through the dielectric. Whenever two conductors face each other across an insulator they form a capacitor — a device that stores charge and resists sudden voltage change — so our wire actually feels two capacitances.

WHY it matters. To flip a wire from "0" to "1" you must first charge up these capacitors. The bigger the total capacitance, the longer that takes → slower chip. So capacitance is the second enemy of speed. In modern packed layouts the sideways line-to-line term usually dominates, but the downward line-to-substrate term is often comparable and must never be dropped.

PICTURE. One copper wire with two sets of field lines: sideways to its neighbour (across gap $s$) and downward to the grounded substrate (across height $h$).

---

Step 5 — Multiplying R and C: the delay appears

WHAT. Signal delay through a wire scales with the product $R\times C$ (charging a capacitor $C$ through a resistor $R$ takes a time proportional to $RC$). Let's multiply our two boxed results.

WHY multiply, not add. The resistor and the capacitor act in series on the same signal: current must flow through $R$ to charge $C$. The physics of "charge a $C$ through an $R$" gives a characteristic time we call $\tau$ (Greek "tau"), and this time is a product — that is the tool we reach for.

PICTURE. The RC "filling a bucket through a straw" analogy: $R$ is the thin straw, $C$ is the bucket; time to fill = (straw thinness) × (bucket size) = $R\times C$.

$\tau =RC=\underset{R}{\underbrace{\rho \frac{L}{wt}}}\times \underset{C_{\parallel }}{\underbrace{{\epsilon }_r{\epsilon }_0\frac{tL}{s}}}$

Watch the thickness $t$: it sits in the denominator of $R$ and in the numerator of $C_{\parallel }$. They cancel:

$\tau =RC=\rho {\epsilon }_r{\epsilon }_0\frac{L\cdot L}{ws}=\rho {\epsilon }_r{\epsilon }_0\frac{L^2}{ws}$

---

Step 6 — The payoff: copper vs aluminum, everything else frozen

WHAT. Compare two identical wires — same $L,w,s$, same dielectric, same substrate — differing only in the metal.

WHY this is clean. In the ratio $t_{\text{delay,Cu}}/t_{\text{delay,Al}}$ every geometric term, every dielectric term, and the $0.38$ distributed-line factor are identical top and bottom, so they cancel. Only $\rho$ survives. This is why we bothered to isolate $\rho$ back in Step 2.

PICTURE. Two identical wire pairs, only the fill colour (metal) differs; a bar chart of their relative delay.

$\frac{t_{\text{delay,Cu}}}{t_{\text{delay,Al}}}=\frac{0.38{\rho }_{Cu}{\epsilon }_r{\epsilon }_0\frac{L^2}{ws}}{0.38{\rho }_{Al}{\epsilon }_r{\epsilon }_0\frac{L^2}{ws}}=\frac{{\rho }_{Cu}}{{\rho }_{Al}}=\frac{1.68}{2.65}\approx 0.63$

• everything except $\rho$ ::: cancels, because it's the same wire shape and same $0.38$ factor
• $0.63$ ::: copper's delay is 63% of aluminum's → a 37% speed-up, for free, just by changing metal

That single number, $\approx 0.63$, is the entire commercial reason the industry re-engineered its whole back-end (Back-End-Of-Line (BEOL)) around copper — even though copper can't be etched and forced the invention of damascene + Chemical Mechanical Polishing (CMP).

---

Step 7 — Edge and degenerate cases (don't get ambushed)

WHAT. Formulas break at extremes. Let's check every geometric limit of $\tau \propto L^2/(ws)$ so nothing surprises you.

PICTURE. Six mini-panels, one per limit, each with a one-line verdict.

• $L\to 0$ (a dot-sized wire). $\tau \propto L^2\to 0$: no delay. A wire of zero length is instant. ✓
• $L$ doubles. $\tau$ grows $\times 4$, not $\times 2$ — the $L^2$ bites. This is why designers buffer long lines and route the longest wires on the fattest top metal layers.
• $w\to 0$ (wire vanishes sideways). $\tau =\rho {\epsilon }_r{\epsilon }_0{L}^2/(ws)\to \infty$: an infinitely thin wire has infinite resistance → infinite delay. Real design keeps $w$ above a minimum feature size set by Photolithography.
• $s\to 0$ (neighbours touching). $C_{\parallel }={\epsilon }_r{\epsilon }_{0}tL/s\to \infty$, so $\tau \to \infty$: the wires would short out. The parallel-plate model is only valid while $s$ is a real gap.
• $s\to \infty$ (neighbour infinitely far). $C_{\parallel }={\epsilon }_r{\epsilon }_{0}tL/s\to 0$ — the sideways coupling vanishes. But delay does not hit zero: the line-to-substrate term $C_{\perp }={\epsilon }_r{\epsilon }_{0}wL/h$ remains, so a lone wire still has a floor delay set by the substrate. (This is exactly why Step 4 kept the third conductor.)
• Ignoring the barrier's resistance. Our $\rho$ was pure copper. The thin Ta/TaN barrier and any electromigration-driven voids nudge the effective $\rho$ upward, so real copper wins a bit less than the ideal $0.63$.

---

The one-picture summary

Everything on this page is one chain: shape a wire → resistance from crowding ($R=\rho L/wt$) → neighbour + substrate make capacitors ($C=C_{\parallel }+C_{\perp }$) → charging through $R$ gives a time constant ($\tau =RC$, real delay $\approx 0.38RC$) → thickness cancels leaving $\tau \propto \rho {\epsilon }_r{L}^2/(ws)$ → freeze geometry and the $0.38$ factor, swap metal → ${\rho }_{Cu}/{\rho }_{Al}\approx 0.63$ → 37% faster.

Recall Feynman retelling — the whole walkthrough in plain words

Picture a tiny copper bar in a ditch. Pushing electricity through it is like people walking a corridor: a long, narrow corridor is hard (that's resistance, and copper is a "slippier" corridor than aluminum). Now the bar has company — a neighbour wire beside it and the silicon floor beneath it. Each of those facing pairs is a charge-storing sandwich (a capacitor); to switch the wire on you first have to fill both sandwiches, which takes time. The delay is roughly the corridor's difficulty times the total sandwich size — an $R$ times $C$, which we call $\tau$. Because the wire is really smeared out (resistance and capacitance sprinkled all along it), the honest step delay is about $0.38$ of that product, not the whole thing — but that fixed fraction doesn't change any comparison. Write it all out and the thickness cancels, leaving delay $\propto L^2/(ws)$: long wires are murder, and the only free lever is the metal. Freeze the shape, swap aluminum for copper, and the delay drops to $63\%$ — a $37\%$ speed-up. But copper can't be etched into wires, so we had to carve the glass and inlay the copper instead. That inlaying is the damascene process.

Recall Quick self-test

• Why does $R$ grow with $L$ but shrink with $A$? ::: longer corridor = more distance; wider doorway = less crowding
• Which two capacitors does one wire feel, and to what? ::: line-to-line $C_{\parallel }$ (to the neighbour) and line-to-substrate $C_{\perp }$ (to the silicon floor)
• Why do we multiply $R$ and $C$ instead of adding? ::: current must flow through $R$ to charge $C$; that charging time is $\tau =RC$
• Is a real wire's delay exactly $RC$? ::: no — a distributed RC line gives $\approx 0.38RC$ at the 50% point
• Why does thickness $t$ vanish from the delay? ::: it's in the denominator of $R$ and numerator of $C_{\parallel }$, so it cancels in the product
• What is the copper-vs-aluminum delay ratio, and why only $\rho$ survives? ::: ${\rho }_{Cu}/{\rho }_{Al}\approx 0.63$; identical geometry and the same $0.38$ factor cancel everything else

---

Related: Copper damascene process · RC delay in interconnects · Aluminum metallization (subtractive) · Reactive Ion Etching · Photolithography