Interconnect: Wire Resistance, Capacitance and Repeaters
Model a long wire as a chain of resistors and capacitors. See its delay grow with the square of its length, and cut it with repeaters.
You Will Be Able To
- Compute the resistance of a wire from its sheet resistance, length and width.
- Explain why the delay of an unbroken wire grows with the square of its length.
- Choose the number and the size of repeaters for a long wire.
- Describe how coupling capacitance makes a quiet wire glitch.
- Lesson
- G9.8
- Time
- 30 min
- Runs on
- ngspice
- Check
- 5 questions
An 8 mm wire, driven by a strong inverter, took 5337.4 ps. The same wire cut into eight pieces by seven inverters took 1828.7 ps. Adding gates made it 2.9 times faster.
- no repeater
- a repeater every 1 mm
Read the plot as a table
| Wire length | no repeater | a repeater every 1 mm |
|---|---|---|
| 1 mm | 163.4 ps | 163.4 ps |
| 2 mm | 450.5 ps | 399 ps |
| 3 mm | 964.2 ps | 637.2 ps |
| 4 mm | 1.478 ns | 875.5 ps |
| 5 mm | 2.443 ns | 1.114 ns |
| 6 mm | 3.408 ns | 1.352 ns |
| 7 mm | 4.372 ns | 1.59 ns |
| 8 mm | 5.337 ns | 1.829 ns |
Measured with ngspice 42 on SKY130, 1.8 V: the mean of both input edges, from the input to the wire end.
A Wire Is a Long Resistor and a Long Capacitor
A wire on a chip is a thin strip of metal. Its resistance is set by the sheet resistance of the layer: R = R_s × length / width. R_s is in ohms per square, the resistance of any square piece of the layer.
| SKY130 layer | Sheet resistance |
|---|---|
| local interconnect (li1) | 12.8 Ω per square |
| metal 1 and metal 2 | 0.125 Ω per square |
| metal 3 and metal 4 | 0.047 Ω per square |
| metal 5 | 0.029 Ω per square |
From the extraction rules shipped with the SKY130 PDK. Higher metal layers are thicker, so their sheet resistance is lower.
A 1 mm metal-1 wire at its minimum width of 0.14 µm is 1000 / 0.14 = 7143 squares long. Its resistance is 893 Ω. That is close to 1 Ω per µm, the value this lesson uses.
- Interconnect
- Neutral
- Structure
- To the layer below
The bottom face of the wire and its edges form a capacitor with the layer below. Wider wires have more.
- To the neighbours
The side walls face the wires next to it. As wires get taller and closer, this coupling becomes the largest part.
- In total
A common estimate for a narrow wire is about 0.2 fF per µm, all parts together. This lesson uses that round value.
Read the diagram as text
- Neighbour wire: same layer
- The wire: about 0.2 fF per µm in all
- Neighbour wire: same layer
- Layer below: substrate or a lower metal
- The wire → Neighbour wire: coupling
- The wire → Neighbour wire: coupling
- The wire → Layer below: area and fringe
The lesson's wire: 1 Ω and 0.2 fF per µm, in 100 µm sections of 100 Ω and 20 fF. These are round teaching values, not extracted ones.
Delay Grows with the Square of the Length
Doubling a wire doubles its resistance and doubles its capacitance. The RC product of the wire grows four times.
| Length | Measured delay | Ratio to the length before | Elmore sum of the wire alone |
|---|---|---|---|
| 1 mm | 163.4 ps | 110 ps | |
| 2 mm | 450.5 ps | 2.76 | 420 ps |
| 4 mm | 1477.8 ps | 3.28 | 1640 ps |
| 8 mm | 5337.4 ps | 3.61 | 6480 ps |
Measured: an inverter of size 8 drives the wire, and another of size 8 receives it. The Elmore sum counts only the wire's own sections.
The ratio climbs towards 4: 2.76, then 3.28, then 3.61. For a short wire the driver's own resistance is a large part of the delay. For a long one the wire's own RC takes over, and that part grows with the length squared.
Repeaters Turn the Square into a Straight Line
Cut the wire into k pieces and put an inverter, a repeater, at the start of each piece after the first. Each piece has 1/k of the length, so its wire delay is 1/k² of the whole. There are k pieces, so the wire part falls to 1/k. Each repeater adds a gate delay.
- v(in)
- v(aend), no repeater
- v(bend), three repeaters
- Time 225 psv(in) = 900 mV · v(aend), no repeater = 1.8 V · v(bend), three repeaters = 345.9 nV
The input rises through 0.9 V.
- Time 1.105 nsv(in) = 1.8 V · v(aend), no repeater = 1.281 V · v(bend), three repeaters = 900 mV
With three repeaters the far end crosses 0.9 V after 880.3 ps. Its edge is sharp: the last repeater drives only 1 mm.
- Time 1.651 nsv(in) = 1.8 V · v(aend), no repeater = 900 mV · v(bend), three repeaters = 1.769 V
With no repeater it takes 1426.2 ps, and the edge is slow. The driver must charge all 4 mm through the wire's own resistance.
Read the plot as a table
| Time | v(in) | v(aend), no repeater | v(bend), three repeaters |
|---|---|---|---|
| 0 s | 0 V | 1.8 V | 345.9 nV |
| 225 ps | 900 mV | 1.8 V | 345.9 nV |
| 500 ps | 1.8 V | 1.773 V | 5.883 µV |
| 1 ns | 1.8 V | 1.371 V | 373 mV |
| 1.105 ns | 1.8 V | 1.281 V | 900 mV |
| 1.5 ns | 1.8 V | 989.7 mV | 1.719 V |
| 1.651 ns | 1.8 V | 900 mV | 1.769 V |
| 2 ns | 1.8 V | 721.7 mV | 1.797 V |
| 2.5 ns | 1.8 V | 514.9 mV | 1.8 V |
| 3 ns | 1.8 V | 368.4 mV | 1.8 V |
Measured with ngspice 42 on SKY130, 1.8 V. The circuits are those of the lab.
| 4 mm wire | Pieces | Measured delay |
|---|---|---|
| no repeater | 1 of 4 mm | 1477.8 ps |
| 1 repeater, size 8 | 2 of 2 mm | 1069.7 ps |
| 3 repeaters, size 8 | 4 of 1 mm | 875.5 ps |
| 7 repeaters, size 8 | 8 of 0.5 mm | 833.0 ps |
| 3 repeaters, size 1 | 4 of 1 mm | 2678.0 ps |
| 3 repeaters, size 4 | 4 of 1 mm | 1106.6 ps |
| 3 repeaters, size 16 | 4 of 1 mm | 812.9 ps |
Measured: mean of both input edges. The driver and the receiver are inverters of size 8 in every row.
The count has an optimum. Going from one to three repeaters saves 194.3 ps. Going from three to seven saves only 42.5 ps, for four more gates.
The size matters as much. Repeaters of size 1 are too weak for 1 mm of wire: 2678.0 ps, worse than no repeater. Size 16 is faster than size 8. But each repeater is twice as large, and it loads the piece before it twice as much.
With a repeater every millimetre, the delay grows almost in a straight line. It is 399.0 ps for 2 mm, 875.5 ps for 4 mm and 1828.7 ps for 8 mm.
Coupling Capacitance and Crosstalk
Part of a wire's capacitance goes to its neighbours, not to ground. When a neighbour switches, it pushes charge through that capacitance into the wire.
- v(a10), aggressor far end
- v(v10), victim far end
- Time 277.7 psv(a10), aggressor far end = 65.07 mV · v(v10), victim far end = 50.07 mV
The aggressor starts to rise, and the victim rises with it.
- Time 423.3 psv(a10), aggressor far end = 900 mV · v(v10), victim far end = 315.9 mV
The aggressor crosses 0.9 V. The victim peaks at 341 mV, then its weak driver slowly pulls it back to 0 V.
Read the plot as a table
| Time | v(a10), aggressor far end | v(v10), victim far end |
|---|---|---|
| 0 s | 345.9 nV | 345.9 nV |
| 200 ps | 345.9 nV | 345.9 nV |
| 277.7 ps | 65.07 mV | 50.07 mV |
| 400 ps | 788.1 mV | 298.2 mV |
| 423.3 ps | 900 mV | 315.9 mV |
| 600 ps | 1.425 V | 327.8 mV |
| 800 ps | 1.646 V | 248.2 mV |
| 1 ns | 1.727 V | 166.8 mV |
| 1.2 ns | 1.761 V | 106.4 mV |
| 1.4 ns | 1.777 V | 66.29 mV |
Measured with ngspice 42 on SKY130, 1.8 V. The coupling value is a round teaching number.
The glitch depends on how hard the victim is held. With an inverter of size 8 holding it, the peak falls from 341 mV to 241 mV.
If a glitch reaches the threshold of a receiver, it can flip a latch. Coupling also changes delay: a neighbour that switches the other way at the same time makes the wire slower.
Asked in interviews
Why does the delay of a long wire grow with the square of its length?
Both its resistance and its capacitance grow with the length. The wire charges its own capacitance through its own resistance, so its RC grows with the length squared.
How do repeaters help?
They cut the wire into k pieces. Each piece has 1/k² of the RC, so the wire part of the delay falls to 1/k. The repeaters add k gate delays, so there is a best count and a best size.
What is crosstalk?
Charge coupled from a switching wire into its neighbour through the capacitance between them. It can glitch a quiet wire or change the delay of a switching one.
Try It
a_up and a_down are the two delays of the bare 4 mm wire: 1426.2 and 1529.4 ps. b_up and b_down are the same wire with three repeaters: 880.3 and 870.7 ps.
Change the three m=8 of the repeaters bR1, bR2 and bR3 to m=1 and run. The repeated wire becomes slower than the bare one: the mean is 2678.0 ps.
Try sizes 4 and 16 too, and compare with the table. Each size trades the delay of a wire piece against the load on the piece before it.
Run the netlist to draw its plots here.
a_up = 1.426228e-09 targ= 1.651228e-09 trig= 2.250000e-10 a_down = 1.529449e-09 targ= 1.170445e-08 trig= 1.017500e-08 b_up = 8.802941e-10 targ= 1.105294e-09 trig= 2.250000e-10 b_down = 8.706728e-10 targ= 1.104567e-08 trig= 1.017500e-08Check Yourself
Common Mistakes
Wrong Make the driver larger to speed up a long wire.
A larger driver only lowers the resistance at the start. The wire's own resistance still charges the wire's own capacitance.
Fix Cut the wire with repeaters. Each piece then has a quarter of the RC for half the length.
Wrong Put a repeater every 100 µm, to be safe.
Every repeater adds its own delay. Seven repeaters on 4 mm gave 833.0 ps against 875.5 ps for three: a small gain for more than twice the repeaters.
Fix Find the spacing where the delay stops falling, and stop there.
Wrong Use minimum-size inverters as repeaters.
A unit inverter cannot drive 1 mm of wire. Three of them on 4 mm gave 2678.0 ps, slower than no repeater at all (1477.8 ps).
Fix Size the repeater for the wire piece it drives: here about 8 unit inverters.
Wrong Treat a quiet wire as safe from noise.
Its neighbour couples charge into it. A quiet 1 mm victim, held low by a unit inverter, rose to 341 mV.
Fix Hold quiet wires with a strong enough driver, space them out, or shield them.
Remember
- Wire resistance: R = R_s × length / width. SKY130 metal 1 is 0.125 Ω per square.
- Wire capacitance goes to the layer below and to the neighbours; about 0.2 fF per µm is a common estimate.
- An unbroken wire's own delay grows with the square of its length.
- Repeaters make the delay grow in a straight line. Both their count and their size have an optimum.
- Coupling capacitance makes quiet wires glitch and changes the delay of switching ones.