Delay: The RC Model, Elmore Delay and Input Slope
Estimate the delay of a gate and of a wire as a resistance charging capacitance, and check each estimate against a measurement.
You Will Be Able To
- Estimate a delay as 0.69 times a resistance times a capacitance.
- Predict how the delay of a gate changes with its load, its size and its input edge.
- Compute the Elmore delay of an RC ladder and of an RC tree.
- State where the Elmore estimate is good and where it is poor.
- Lesson
- G9.6
- Time
- 30 min
- Runs on
- ngspice
- Check
- 5 questions
A gate met its timing in the block-level run with a 10 fF load: 39.2 ps. On the chip its output also drove a long wire, and the load was 40 fF. The gate now took 99.5 ps. Nothing in the gate had changed.
- v(in)
- v(y2), 10 fF
- v(y3), 20 fF
- v(y4), 40 fF
- Time 125 psv(in) = 900 mV · v(y2), 10 fF = 1.877 V · v(y3), 20 fF = 1.849 V · v(y4), 40 fF = 1.828 V
The input crosses 0.9 V. Each delay is counted from here to the same level at the output.
- Time 164.2 psv(in) = 1.8 V · v(y2), 10 fF = 900.4 mV · v(y3), 20 fF = 1.309 V · v(y4), 40 fF = 1.543 V
With 10 fF the output crosses 0.9 V after 39.2 ps.
- Time 184.4 psv(in) = 1.8 V · v(y2), 10 fF = 263.5 mV · v(y3), 20 fF = 900.1 mV · v(y4), 40 fF = 1.322 V
With 20 fF it takes 59.4 ps. Ten more femtofarads cost 20.2 ps.
- Time 224.5 psv(in) = 1.8 V · v(y2), 10 fF = 5.719 mV · v(y3), 20 fF = 231.2 mV · v(y4), 40 fF = 900 mV
With 40 fF it takes 99.5 ps. Twenty more cost 40.1 ps: twice as much.
Read the plot as a table
| Time | v(in) | v(y2), 10 fF | v(y3), 20 fF | v(y4), 40 fF |
|---|---|---|---|---|
| 0 s | 0 V | 1.8 V | 1.8 V | 1.8 V |
| 125 ps | 900 mV | 1.877 V | 1.849 V | 1.828 V |
| 164.2 ps | 1.8 V | 900.4 mV | 1.309 V | 1.543 V |
| 184.4 ps | 1.8 V | 263.5 mV | 900.1 mV | 1.322 V |
| 200 ps | 1.8 V | 64.3 mV | 603.8 mV | 1.155 V |
| 224.5 ps | 1.8 V | 5.719 mV | 231.2 mV | 900 mV |
| 400 ps | 1.8 V | 346.2 nV | 17.15 µV | 15.41 mV |
| 600 ps | 1.8 V | 345.9 nV | 346.2 nV | 46.62 µV |
| 800 ps | 0 V | 1.752 V | 1.429 V | 862.3 mV |
| 1 ns | 0 V | 1.8 V | 1.797 V | 1.688 V |
| 1.2 ns | 0 V | 1.8 V | 1.8 V | 1.792 V |
Measured with ngspice 42 on SKY130, 1.8 V.
A Resistor Charging a Capacitor
A switching gate is a transistor that is on, between a supply rail and the output. The output is a capacitor. So the simplest model of delay has two parts: one resistance R and one capacitance C.
After a step, the capacitor voltage follows an exponential curve. It reaches half of its final value after 0.69 × R × C. The product RC is the time constant.
| Circuit | RC | 0.69 × RC | Measured |
|---|---|---|---|
| 1 kΩ into 40 fF | 40 ps | 27.6 ps | 27.7 ps |
The first circuit of the lab. The delay is from 50 % of the step to 50 % at the capacitor.
For one R and one C the rule is exact. A real gate is not one resistor, and a real wire is not one capacitor. The rest of the lesson finds an R and a C that still give a good answer.
Gate Delay Grows in a Straight Line with the Load
| Load | Output falls | Output rises |
|---|---|---|
| none | 15.9 ps | 18.4 ps |
| 5 fF | 28.9 ps | 34.2 ps |
| 10 fF | 39.2 ps | 48.2 ps |
| 20 fF | 59.4 ps | 75.7 ps |
| 40 fF | 99.5 ps | 130.4 ps |
Measured: the unit inverter, 50 ps input edge, 50 % to 50 %. Every drain and source has the junction of a contacted diffusion, 0.29 µm long, as in a laid-out cell.
From 10 fF to 40 fF the fall delay grows by 60.3 ps. In the model that is 0.69 × R × 30 fF, so R is 2.91 kΩ. This is the effective resistance of the 1 µm NMOS. The 2 µm PMOS gives 3.97 kΩ.
The delay with no load is not zero: 15.9 ps. The drains of the two transistors are capacitors as well: the junctions of their diffusions, and the overlap with their gates. This part is the diffusion capacitance, and it is always there.
Size works the other way. An inverter twice as wide has half the resistance. Into 20 fF it falls in 37.7 ps. The unit inverter into 10 fF takes 39.2 ps. Delay follows the ratio of load to size.
- Device
- Load
- I/O
- Power
- No load
Only the gate's own drains must be charged. The inverter still needs 15.9 ps.
- With a load
The same current now charges the load too. Each extra 10 fF adds about 20.1 ps to the fall.
- A wider gate
Twice the width halves R. It also doubles the gate's own drain capacitance and its input capacitance.
Read the diagram as text
- Transistor that is on: effective resistance R
- Own drains: diffusion capacitance
- Output node: crosses 50 % after 0.69 RC
- Load: gates and wire
- Transistor that is on → Output node: current
- Output node → Own drains: charges
- Output node → Load: charges
Delay ≈ 0.69 × R × (diffusion capacitance + load).
A Slow Input Edge Makes a Slow Gate
The model assumed that the transistor turns on at once. With a slow input edge it turns on gradually. For part of the edge it is only half on, and the other transistor is still fighting it.
- v(in1), 50 ps edge
- v(y1)
- v(in4), 400 ps edge
- v(y4)
- Time 125 psv(in1), 50 ps edge = 900 mV · v(y1) = 1.877 V · v(in4), 400 ps edge = 112.5 mV · v(y4) = 1.809 V
The fast input crosses 0.9 V.
- Time 164.2 psv(in1), 50 ps edge = 1.8 V · v(y1) = 900.3 mV · v(in4), 400 ps edge = 288.9 mV · v(y4) = 1.813 V
Its output follows 39.2 ps later.
- Time 300 psv(in1), 50 ps edge = 1.8 V · v(y1) = 3.2 µV · v(in4), 400 ps edge = 900 mV · v(y4) = 1.769 V
The slow input crosses 0.9 V. The output has already begun to fall, because the NMOS turned on earlier in the edge.
- Time 390 psv(in1), 50 ps edge = 1.8 V · v(y1) = 346.3 nV · v(in4), 400 ps edge = 1.305 V · v(y4) = 899.9 mV
The output of the slow input crosses 0.9 V 90.0 ps after its input.
Read the plot as a table
| Time | v(in1), 50 ps edge | v(y1) | v(in4), 400 ps edge | v(y4) |
|---|---|---|---|---|
| 0 s | 0 V | 1.8 V | 0 V | 1.8 V |
| 125 ps | 900 mV | 1.877 V | 112.5 mV | 1.809 V |
| 164.2 ps | 1.8 V | 900.3 mV | 288.9 mV | 1.813 V |
| 200 ps | 1.8 V | 64.27 mV | 450 mV | 1.816 V |
| 300 ps | 1.8 V | 3.2 µV | 900 mV | 1.769 V |
| 390 ps | 1.8 V | 346.3 nV | 1.305 V | 899.9 mV |
| 400 ps | 1.8 V | 346.1 nV | 1.35 V | 732.1 mV |
| 600 ps | 1.8 V | 345.9 nV | 1.8 V | 522.5 nV |
| 800 ps | 9.643 mV | 261.4 mV | 1.8 V | 345.9 nV |
| 1 ns | 0 V | 1.799 V | 1.8 V | 345.9 nV |
| 1.2 ns | 0 V | 1.8 V | 1.35 V | -5.506 mV |
| 1.4 ns | 0 V | 1.8 V | 450 mV | 743.2 mV |
| 1.6 ns | 0 V | 1.8 V | 0 V | 1.799 V |
Measured with ngspice 42 on SKY130, 1.8 V. Only the falling outputs are marked.
| Input edge (0 to 100 %) | Output falls | Output rises |
|---|---|---|
| 10 ps | 30.6 ps | 39.3 ps |
| 50 ps | 39.2 ps | 48.2 ps |
| 100 ps | 50.4 ps | 60.0 ps |
| 200 ps | 68.0 ps | 82.3 ps |
| 400 ps | 90.0 ps | 112.8 ps |
Measured: the unit inverter into 10 fF.
A slow edge at one gate comes from a heavy load on the gate before it. So a large load costs twice: once in its own gate and once in the next. Timing tools store gate delay in tables of input edge against load for this reason.
Elmore Delay: Many Resistors and Capacitors
A wire has resistance and capacitance all along its length. A series stack of transistors has a small capacitor at every inner node. One RC product is not enough for these.
The Elmore delay is a sum. Take each capacitor in turn. Multiply it by the resistance between the source and that capacitor. Count only resistors that are also on the path to your node. Add the products.
- Source
- Load
- Data
- The capacitor at n1
It is charged through 1 kΩ: 1 kΩ × 10 fF = 10 ps. Sum so far: 10 ps.
- The capacitor at n2
It is charged through 2 kΩ: 20 ps. Sum so far: 30 ps.
- The capacitor at n3
It is charged through 3 kΩ: 30 ps. Sum so far: 60 ps.
- The capacitor at n4
It is charged through all 4 kΩ: 40 ps. The Elmore delay of n4 is 100 ps, and 0.69 × 100 ps = 69 ps. Measured: 75.3 ps.
Read the diagram as text
- Source: ideal step
- n1: 10 fF behind 1 kΩ
- n2: 10 fF behind 2 kΩ
- n3: 10 fF behind 3 kΩ
- n4: 10 fF behind 4 kΩ
- Source → n1: 1 kΩ
- n1 → n2: 1 kΩ
- n2 → n3: 1 kΩ
- n3 → n4: 1 kΩ
Lumping everything at the end would give 4 kΩ × 40 fF = 160 ps. Spreading the capacitance along the wire makes it faster.
| Node | Elmore sum | 0.69 × Elmore | Measured |
|---|---|---|---|
| n1 | 40 ps | 27.6 ps | 11.2 ps |
| n2 | 70 ps | 48.3 ps | 40.9 ps |
| n3 | 90 ps | 62.1 ps | 64.7 ps |
| n4 | 100 ps | 69.0 ps | 75.3 ps |
For n1 the sum counts all four capacitors behind the first resistor: 1 kΩ × 40 fF. For n2 it adds 1 kΩ × 30 fF, and so on.
The estimate is good at the far end: 69 ps against 75.3 ps. It is poor near the source. Node n1 crosses 50 % after 11.2 ps, far sooner than 27.6 ps. Use Elmore for the end of a path, where the delay matters.
A series stack is a short ladder. In a three-input NAND the fall takes 34.2 ps when the switching input is at the top of the stack. It takes 47.9 ps when that input is at the bottom. Then the two inner nodes of the stack must be emptied as well.
Asked in interviews
What is the Elmore delay?
An estimate of the delay of an RC tree. Take each capacitor and the path from the source to it. Multiply the capacitor by the resistance this path shares with the path to your node. The sum is the Elmore delay. Multiply by 0.69 for the 50 % point.
Why does delay depend on the input slope?
A slow edge turns the driving transistor on gradually, so it delivers less current at first. The opposing transistor also stays on for longer. Both make the output move later than with a fast edge.
Why is a distributed wire faster than the same R and C lumped together?
In a ladder, capacitance near the driver is charged through only a small part of the resistance. The Elmore sum for a uniform line is about half of the total R times the total C.
Try It
t_lumped is the single RC: 27.7 ps, which is 0.69 × 40 ps. t_n4 is the far end of the ladder: 75.3 ps.
The ladder has the same total resistance as four of the lumped resistors, and the same total capacitance as the lumped capacitor. Its far end is still much faster than 0.69 × 160 ps = 110 ps.
Work out the Elmore sum for n4 after doubling R4 to 2 kΩ: 10 + 20 + 30 + 50 = 110 ps. Then change R4 and run. Compare t_n4 with 0.69 × 110 ps.
Run the netlist to draw its plots here.
t_lumped = 2.772571e-11 targ= 1.282257e-10 trig= 1.005000e-10 t_n1 = 1.117983e-11 targ= 1.116798e-10 trig= 1.005000e-10 t_n2 = 4.093309e-11 targ= 1.414331e-10 trig= 1.005000e-10 t_n3 = 6.472133e-11 targ= 1.652213e-10 trig= 1.005000e-10 t_n4 = 7.534840e-11 targ= 1.758484e-10 trig= 1.005000e-10Check Yourself
Common Mistakes
Wrong Add up all the capacitance, multiply by all the resistance, and call it the delay.
For the ladder of the lab that gives 4 kΩ × 40 fF = 160 ps. The Elmore sum is 100 ps and the measured delay is 75.3 ps.
Fix Multiply each capacitor by the resistance between the source and that capacitor only, then add.
Wrong In a tree, add the resistance of a side branch to the path.
Charge for a side branch flows only through the shared part of the path. The branch resistor itself is not on the way to your node.
Fix For a capacitor on a side branch, use the resistance the two paths share.
Wrong Forget the capacitance of the gate's own transistors.
The unit inverter with no load at all still takes 15.9 ps to fall. Its drains are a load too.
Fix Add the diffusion capacitance at the output to the load before you multiply.
Wrong Quote a gate delay without the input edge.
After a 10 ps edge the inverter falls in 30.6 ps. After a 400 ps edge, with the same load, it takes 90.0 ps.
Fix State the input edge with every delay, or use a table of delay against edge and load.
Remember
- Delay ≈ 0.69 × R × C: the time for a capacitor to reach half of a step.
- Gate delay grows in a straight line with the load. The gate's own drains add a fixed part.
- A wider gate has less resistance. Delay follows the ratio of load to size.
- A slow input edge adds delay, so a heavy load also slows the next gate.
- Elmore delay: each capacitor times the resistance shared with the path, summed.