IAE Academy

FOR CORE AND IT JOBS · CHAPTER 3 OF 7

Delay, power and scaling

A gate switches by charging or discharging a capacitance through a resistance. This chapter turns that picture into delay, buffer sizing, power and the scaling rules, so you can estimate how fast a circuit is and how much power it draws.

Updated 30 Oct 2026

What you will be able to do

You will model a switching gate as a resistor charging a capacitor and estimate its delay with and, for a chain, the Elmore sum. You will size a buffer chain for a large load or a long wire. You will derive dynamic power and judge what a lower supply costs in speed. You will say what scaling gave and why it stopped.

Before you start

  1. Chapter 2, Static CMOS logic gates

  2. Charge and energy on a capacitor: and stored energy

  3. The exponential charging of a capacitor through a resistor

  4. Natural logarithms, integrating a decaying exponential, and differentiating a quotient

  5. The square law in saturation, , with and

Delay: a gate as a resistor charging a capacitor

When an input changes, the network of the gate that turns on behaves like a resistor . The output node, with its wire and the inputs of the gates it drives, behaves like a capacitor to ground. Switching is charging through the pull-up or discharging through the pull-down. A wider transistor has a smaller and a larger input capacitance, and both facts are used below.

PROPAGATION DELAY OF ONE RC STAGE

Holds for constant , an ideal input step at and a rising output starting at 0; a falling output, , gives the same . is the supply. is the time to reach : setting gives . kΩ times fF gives ps.

For a chain, a source drives to node 1, which has to ground; runs to node 2 with , and to node 3 with . All capacitors start empty and the source steps to . Each capacitor ends at , and all charge that ends on and passes through , so carries . Likewise carries and carries .

The drop across is times its current, and the time integral of a current is the charge it carries. So each drop, integrated over time, is times the charge above. is the sum of the three drops, so its time integral is the sum of the three products, in volt-seconds. Dividing by leaves seconds: a delay.

ELMORE DELAY OF A THREE-STAGE RC CHAIN

Holds for a step input, empty capacitors and linear resistors, in the netlist above; is the delay at node 3, in ps for kΩ and fF. One stage gives , since , so . Node 3 does not rise as one exponential, so this is an estimate; its merit is that needs no solving of the circuit.

WATCH OUT FOR

Each multiplies all the capacitance downstream of it, not upstream. Test: the last resistor multiplies only the last capacitor.

WORKED EXAMPLE

A source steps to and drives the three-stage chain above with kΩ and fF. Find the Elmore delay at node 3 and an estimate of .

Solution
  1. feeds all three capacitors, since every capacitor is charged through it: 2 kΩ × 150 fF = 300 ps = 0.30 ns.

  2. feeds and : 2 kΩ × 100 fF = 0.20 ns.

  3. feeds only : 2 kΩ × 50 fF = 0.10 ns.

  4. Add the three terms: = 0.60 ns.

  5. The Elmore sum plays the part of in the one-stage result, so multiply by 0.69: 0.69 × 0.60 ns ≈ 0.41 ns.

Answer: = 0.60 ns and 0.41 ns.

Driving a large load

Let a unit inverter have on-resistance , input capacitance and its own output capacitance , where the self-load ratio is about 1. Scaled by , its resistance becomes and both capacitances grow by . Let it drive a load times its own input capacitance, . This ratio is the fan-out: a gate driving copies of itself has . Dropping the factor 0.69 that every stage shares, its delay is , with . The size cancels: the delay depends only on .

A wire's capacitance grows with its length, so a long wire, like a large load, gives the gate driving it a huge and a long delay. Instead drive it through a chain of inverters, the first being the original small one, each times larger than the one before. The overall ratio , with the input capacitance of the first inverter, equals , so . The total delay is . The load fixes . Neglecting for the outline, only is left to minimise.

BEST STAGE RATIO OF A BUFFER CHAIN

Holds with neglected. With , the same step on gives , so , where . At it is 3.64 and at it is 3.73, so from 3 to 5 stays within 4% of the best delay. Round to a whole number.

WATCH OUT FOR

More stages are not always faster, because each inverter adds its own . Test: compare the total delay for and .

WORKED EXAMPLE

A load is 81 times the input capacitance of the first inverter. Take each stage delay as with = 10 ps. Compare driving the load with that one inverter against a chain of ratio 3.

Solution
  1. Directly, = 81, so the delay is 10 ps × (1 + 81) = 820 ps, since one small gate must charge the whole load.

  2. For ratio 3, = 81 gives = 4 stages.

  3. Each stage delay is 10 ps × (1 + 3) = 40 ps, because each stage drives 3 times its own input capacitance.

  4. Four stages in series: 4 × 40 ps = 160 ps.

  5. The single inverter inverts, and keeping that needs an odd number of stages. Three stages of ratio 4.33 give 3 × 10 ps × 5.33 ≈ 160 ps, as fast as four, because the best , 3.4, lies between 3 and 4. Far from it the delay grows: one stage took 820 ps.

Answer: Direct drive takes 820 ps. The four-stage chain takes 160 ps, about five times faster.

Power and scaling

Charge a load from 0 to : the supply delivers charge at voltage , so it gives energy . Only ends up stored on ; the other half becomes heat in the pull-up, whatever its resistance. On discharge the stored half becomes heat in the pull-down. So each 0 to 1 transition of the output, with the 1 to 0 that follows, costs from the supply.

DYNAMIC POWER

Holds for an output that swings fully between 0 and . is the probability that the output makes a 0 to 1 transition in a clock cycle, the switched capacitance, the clock frequency; with in F, in V and in Hz, is in W. The on-resistance does not appear.

Two more terms exist. Short-circuit power is drawn while an input lies between and , when both networks conduct (possible only if ); a slow input edge makes it larger. Leakage power, with the supply current when nothing switches, is drawn even by an idle circuit. One source is the subthreshold current: a transistor that is off, with below its threshold, still passes a small current, and that current grows exponentially as the threshold is lowered.

GATE DELAY AGAINST SUPPLY VOLTAGE

An estimate in the square-law model with : a constant current moves charge . and belong to the switching transistor: and , or and . The table takes = 0.3 V; energy is and delay is over its value at 1.2 V.

V_DD (V)

Energy per switch, relative

Delay, relative

1.2

1.00

1.00

0.9

0.56

1.69

0.6

0.25

4.50

In the table, halving from 1.2 V to 0.6 V saves three quarters of the energy but makes the gate 4.5 times slower. Constant-field scaling multiplies lengths, widths and the oxide thickness by , where is the scaling factor, and and by too. With mobility unchanged, is multiplied by , and so is . The current is multiplied by , and gate capacitance by . So delay is multiplied by and energy per switch by .

CONSTANT-FIELD SCALING OF DELAY AND POWER

Holds when the clock rises by the factor with the faster gate and is unchanged. is the dynamic power of one gate. Area per gate, , is multiplied by , so power per unit area, over that area, is unchanged.

This stopped because could not keep falling. At room temperature, the subthreshold current rises about tenfold for every 60 to 100 mV taken off , and that slope does not shrink with the transistor, so leakage would have grown exponentially. , and with it , almost stopped falling. With fixed and multiplied by , a clock raised by leaves unchanged, so power per unit area is multiplied by . A clock held fixed gives on an area as large: still times the power per unit area. The heat a chip can shed now limits its clock.

Check yourself

CHECK YOURSELF 1

A 20 fF node swings fully between 0 and . It makes a 0 to 1 transition in 10% of the clock cycles. = 1 V and the clock is 1 GHz. Find the dynamic power in µW.

Answer and solution

Answer: (accepted range )

  1. .

  2. = 0.1 × 20 fF = 2 fF.

  3. Multiply by = 1 V² and = Hz: W.

  4. = 2 µW.

CHECK YOURSELF 2

A step source drives = 1 kΩ to node 1, which has = 20 fF to ground; = 3 kΩ runs from node 1 to node 2, which has = 10 fF to ground. Both capacitors start empty. Find the Elmore delay at node 2 in ps.

Answer and solution

Answer: (accepted range )

  1. feeds both capacitors: 1 kΩ × (20 + 10) fF = 30 ps.

  2. feeds only : 3 kΩ × 10 fF = 30 ps.

  3. Sum: 60 ps.

CHECK YOURSELF 3

With = 0.3 V, the supply is lowered from 1.2 V to 0.75 V. In the square-law delay model of this chapter, by what factor does the gate delay change?

  1. (A)

    1.6 times longer

  2. (B)

    2.5 times longer

  3. (C)

    4 times longer

  4. (D)

    6.25 times longer

Answer and solution

Answer: (B)

  1. is proportional to .

  2. At 1.2 V: 1.2 / 0.9² = 1.481.

  3. At 0.75 V: 0.75 / 0.45² = 3.704.

  4. Ratio 3.704 / 1.481 = 2.5.

CHECK YOURSELF 4

In a buffer chain, the input capacitance of each inverter is 4 times that of the one before, and the load is 4 times the input capacitance of the last inverter. The load is 256 times the input capacitance of the first inverter. How many inverters are there?

  1. (A)

    3

  2. (B)

    4

  3. (C)

    5

  4. (D)

    8

Answer and solution

Answer: (B)

  1. Each inverter multiplies the capacitance by 4 and the load adds one more factor of 4, so with inverters.

  2. 256 = , and = 256, so = 4.

Used in these GATE topics