Logical Effort and Sizing a Path
Size the gates of a path for the least delay with logical effort, and choose how many stages a large load needs.
You Will Be Able To
- Write the delay of a stage as g·h + p and say what each letter stands for.
- Compute the path effort F = G·B·H of a path with a branch.
- Size each stage of a path for equal stage effort.
- Choose the number of stages that drives a large load fastest.
- Lesson
- G9.7
- Time
- 35 min
- Runs on
- ngspice
- Check
- 5 questions
A unit inverter had to drive a load of 64 unit inverters. Alone it needed 583.4 ps. Two more inverters between it and the load, each four times the size of the one before, cut that to 185.9 ps. More gates on the path made it faster.
- v(in)
- v(n1), size 1
- v(n2), size 4
- v(n3), size 16
- Time 225 psv(in) = 900 mV · v(n1), size 1 = 1.867 V · v(n2), size 4 = 1.338 mV · v(n3), size 16 = 1.8 V
The input rises through 0.9 V. The first inverter is a unit inverter, so the input sees a small load.
- Time 270.1 psv(in) = 1.8 V · v(n1), size 1 = 900.2 mV · v(n2), size 4 = -29.87 mV · v(n3), size 16 = 1.798 V
The first stage, size 1, drives size 4. It crosses 0.9 V after 45.1 ps. Its input is a sharp 50 ps edge.
- Time 338.3 psv(in) = 1.8 V · v(n1), size 1 = 25.94 mV · v(n2), size 4 = 899.7 mV · v(n3), size 16 = 1.848 V
The second stage, size 4, drives size 16. It takes 68.1 ps more, since its input edge comes from a real gate.
- Time 402.3 psv(in) = 1.8 V · v(n1), size 1 = 23.8 mV · v(n2), size 4 = 1.622 V · v(n3), size 16 = 900.3 mV
The last stage, size 16, drives the load of 64 in 64.1 ps. Both later stages have a fan-out of 4 and take about the same time.
Read the plot as a table
| Time | v(in) | v(n1), size 1 | v(n2), size 4 | v(n3), size 16 |
|---|---|---|---|---|
| 0 s | 0 V | 1.8 V | 345.9 nV | 1.8 V |
| 225 ps | 900 mV | 1.867 V | 1.338 mV | 1.8 V |
| 270.1 ps | 1.8 V | 900.2 mV | -29.87 mV | 1.798 V |
| 338.3 ps | 1.8 V | 25.94 mV | 899.7 mV | 1.848 V |
| 402.3 ps | 1.8 V | 23.8 mV | 1.622 V | 900.3 mV |
| 500 ps | 1.8 V | 4.412 mV | 1.777 V | 8.703 mV |
| 1 ns | 1.8 V | 346 nV | 1.8 V | 345.8 nV |
| 1.5 ns | 1.8 V | 345.9 nV | 1.8 V | 345.9 nV |
| 2 ns | 0 V | 1.786 V | 16.99 mV | 1.255 V |
| 2.5 ns | 0 V | 1.8 V | 351.2 nV | 1.8 V |
| 3 ns | 0 V | 1.8 V | 345.9 nV | 1.8 V |
Measured with ngspice 42 on SKY130, 1.8 V, 50 ps input edge. The run is the lab of this lesson.
Stage Delay Is g·h + p
G9.6 showed that gate delay grows in a straight line with the load. Logical effort writes that line without units: d = g·h + p.
| Letter | Name | Meaning |
|---|---|---|
| h | electrical effort | load capacitance divided by the input capacitance of the gate: its fan-out |
| g | logical effort | how much worse the gate is than an inverter at driving a load; 1 for an inverter |
| p | parasitic delay | the delay of the gate with no load, from its own drains |
| τ | unit of delay | the slope of the inverter line: delay per unit of fan-out |
The delay d is counted in units of τ. The product f = g·h is the stage effort.
- v(h4n1)
- v(h4n2)
- v(h4n3)
- Time 170.1 psv(h4n1) = 900.3 mV · v(h4n2) = -29.07 mV · v(h4n3) = 1.799 V
The first stage falls through 0.9 V.
- Time 241.2 psv(h4n1) = 27.16 mV · v(h4n2) = 899.7 mV · v(h4n3) = 1.845 V
The second stage rises through 0.9 V 71.1 ps later. This is the fan-out-of-4 delay for a rising output.
- Time 309.4 psv(h4n1) = 7.057 mV · v(h4n2) = 1.625 V · v(h4n3) = 900 mV
The third stage falls 68.2 ps after that. Its load is the same four gates, with nothing behind them.
Read the plot as a table
| Time | v(h4n1) | v(h4n2) | v(h4n3) |
|---|---|---|---|
| 0 s | 1.8 V | 345.9 nV | 1.8 V |
| 170.1 ps | 900.3 mV | -29.07 mV | 1.799 V |
| 241.2 ps | 27.16 mV | 899.7 mV | 1.845 V |
| 309.4 ps | 7.057 mV | 1.625 V | 900 mV |
| 500 ps | 24.09 µV | 1.8 V | 29.9 µV |
| 1 ns | 345.8 nV | 1.8 V | 346.6 nV |
| 1.5 ns | 1.8 V | 48.1 µV | 1.8 V |
| 2 ns | 1.8 V | 345.9 nV | 1.8 V |
Measured with ngspice 42 on SKY130, 1.8 V. The same run holds the chains with fan-out 1, 2, 3 and 6.
| Fan-out h | Inverter | NAND2 | NOR2 |
|---|---|---|---|
| 1 | 24.6 ps | 33.7 ps | 41.4 ps |
| 2 | 39.4 ps | 51.3 ps | 63.8 ps |
| 3 | 54.4 ps | 69.6 ps | 86.7 ps |
| 4 | 69.5 ps | 88.0 ps | 109.9 ps |
| 6 | 99.9 ps | 125.2 ps | 156.7 ps |
Measured: the mean of rise and fall of the middle stage. Each gate drives h copies of itself.
The inverter column is a straight line: 15.1 ps for each unit of fan-out, plus 9.4 ps. So τ is 15.1 ps in this process. The delay at a fan-out of 4, the FO4 delay, is 69.5 ps.
Logical Effort Compares a Gate with an Inverter
Size a gate so that it drives as well as the unit inverter. Then compare input capacitances. The ratio is the logical effort g.
| Gate, sized for unit drive | Widths at one input | Input capacitance | g in the model | p in the model |
|---|---|---|---|---|
| Inverter | NMOS 1, PMOS 2 | 3 units | 1 | 1 |
| NAND2 | NMOS 2 (series), PMOS 2 | 4 units | 4/3 | 2 |
| NOR2 | NMOS 1, PMOS 4 (series) | 5 units | 5/3 | 2 |
The model assumes a PMOS needs twice the width of an NMOS, and that capacitance follows width.
The model gives the NOR a larger g than the NAND. It needs wide PMOS in series, so each input sees more capacitance for the same drive.
| Gate | Slope of the measured line | g = slope / τ | p = offset / τ |
|---|---|---|---|
| Inverter | 15.1 ps per unit | 1.00 | 0.62 |
| NAND2 | 18.3 ps per unit | 1.22 | 0.99 |
| NOR2 | 23.1 ps per unit | 1.53 | 1.18 |
Fitted to the measured table above (least squares over the five fan-outs). Every drain and source has the junction of a contacted diffusion, 0.29 µm long.
The measured efforts keep the order of the model: inverter, NAND, NOR. The values differ. The NAND measures 1.22 against 4/3, and the inverter's parasitic delay is 0.62 τ, under the model's 1. These netlists give every transistor the drain junction of a laid-out cell; without it the inverter's p falls to about 0.4. Use the model to choose sizes, and a simulator for the final delay.
Path Effort and Equal Stage Effort
A path is a chain of stages between a fixed input capacitance and a fixed load. Three numbers describe it.
- Datapath
- Load
- Data
- G: multiply the logical efforts
G = 1 × 4/3 × 1 = 4/3.
- B: multiply the branching efforts
The NAND drives three equal gates and one is on the path. Its branching effort is 3. So B = 3.
- H: load over input
H = 16.
- F = G·B·H
F = 4/3 × 3 × 16 = 64. With three stages the best stage effort is the cube root: 4.
Read the diagram as text
- Inverter: g = 1
- NAND2: g = 4/3
- Inverter: g = 1
- Load: 16 times the path input
- Two side branches: copies of the last inverter
- Inverter → NAND2
- NAND2 → Inverter: on the path
- NAND2 → Two side branches: off the path
- Inverter → Load
The path delay is least when every stage carries the same effort f = F^(1/N).
The least delay is N·f + P, where P is the sum of the parasitic delays. Here that is 3 × 4 + 4 = 16 τ.
The sizes follow from the load backwards. Each stage needs an input capacitance of g × b × (the capacitance it drives on the path) / f.
The Best Number of Stages
For a large F, more stages can be faster. Each added inverter lowers the effort per stage but adds its own parasitic delay. The best stage effort is close to 4.
| Stages | Sizes | Measured into 64 unit inverters | Model |
|---|---|---|---|
| 1 | 1 | 583.4 ps | 65.0 τ |
| 2 | 1, 8 | 207.9 ps | 18.0 τ |
| 3 | 1, 4, 16 | 185.9 ps | 15.0 τ |
| 4 | 1, 3, 8, 23 | 194.6 ps | 15.3 τ |
| 5 | 1, 2, 5, 12, 28 | 211.8 ps | 16.5 τ |
Measured: mean of both input edges, 50 ps edge at the first stage. Model: equal stage effort with p = 1 per inverter.
Measurement and model agree on the winner: three stages. They also agree that the curve is flat near the best count. Four stages cost only 8.7 ps more than three.
Sizing matters as much as the count. Three unit inverters in a row, with no growth, need 634.4 ps for the same load.
Asked in interviews
What is logical effort?
A way to write gate delay as d = g·h + p, in units of τ. The letter h is the fan-out. The effort g says how much input capacitance the gate needs for the drive of an inverter. The parasitic delay p is its delay with no load. It lets you size a path by hand.
How do you size a path for the least delay?
Compute F = G·B·H. With N stages, give every stage the effort f = F^(1/N). Work back from the load: each input capacitance is g times b times the capacitance it drives, divided by f.
Why is a fan-out of about 4 used for buffer chains?
The delay of a chain is N·f + N·p. A small f needs many stages and a large f makes each stage slow. The sum is least for f between 3 and 4, and it changes little nearby.
Try It
d_up is the delay from the rising input to the last stage: 177.3 ps. d_down is the same for the falling input: 194.5 ps. Their mean is 185.9 ps.
Change the three m= values of stages B2 and B3 to 1 and run: three unit inverters need 634.4 ps. The last one drives all 64 loads alone.
Delete stages B2 and B3 and connect the load to n1. Change both n3 in the meas lines to n1. One inverter needs 583.4 ps.
Run the netlist to draw its plots here.
d_up = 1.773366e-10 targ= 4.023366e-10 trig= 2.250000e-10 d_down = 1.945235e-10 targ= 1.969523e-09 trig= 1.775000e-09 t_in = 2.250000e-10 t_end = 4.023366e-10 t_n1 = 2.701265e-10 t_n2 = 3.382561e-10Check Yourself
Common Mistakes
Wrong Drive a large load with one very large gate.
The large gate is itself a large load for the gate before it. One unit inverter into 64 took 583.4 ps here. Three stages took 185.9 ps.
Fix Grow the size step by step, by a factor of about 4 per stage.
Wrong Use the logical effort of a NAND for a NOR: 4/3.
A NOR has its PMOS in series, so they must be twice as wide again. Its input capacitance is 5 units against 3 for the inverter.
Fix In the model, a two-input NAND has g = 4/3 and a two-input NOR has g = 5/3.
Wrong Leave the branching effort out of F.
A stage that also feeds side branches must charge them too. Its real load is b times the load on the path.
Fix Multiply by b at every stage whose output splits: F = G·B·H.
Wrong Add stages until each one has the smallest possible effort.
Every stage adds its own parasitic delay. Five stages took 211.8 ps here, more than three.
Fix Aim for a stage effort near 4. The stage count that gets closest wins.
Remember
- Stage delay in units of τ: d = g·h + p.
- Model values: inverter g = 1; NAND2 g = 4/3; NOR2 g = 5/3.
- Path effort F = G·B·H. Include the branching effort wherever an output splits.
- Least delay: every stage has the same effort f = F^(1/N). Size backwards from the load.
- The best stage effort is near 4: one stage per factor of 4 in F.