The Transimpedance Amplifier
Turn a photodiode current into a voltage with shunt-shunt feedback, then set the bandwidth, the peaking and the noise by design.
You Will Be Able To
- Explain why the feedback resistor sets the transimpedance and the loop sets the bandwidth.
- Predict the peak that the photodiode capacitance causes, and choose CF to remove it.
- Measure the input-referred noise current and compare it with the noise of RF.
- Compare an op-amp TIA with a one-inverter TIA on the same photodiode.
- Lesson
- E10.3
- Time
- 35 min
- Runs on
- ngspice
- Check
- 5 questions
An optical receiver put 50 kΩ around a 100 MHz op-amp, with a 1 pF photodiode at the input. The response peaked by 14.7 dB at 17.8 MHz, and every light pulse rang.
- no CF
- CF = 0.25 pF
- CF = 1 pF
- Frequency 100 kHzno CF = 93.97 dB · CF = 0.25 pF = 93.97 dB · CF = 1 pF = 93.97 dB
At low frequency all three give 93.97 dBΩ. That is 49.95 kΩ, set by RF.
- Frequency 3.269 MHzno CF = 94.26 dB · CF = 0.25 pF = 93.96 dB · CF = 1 pF = 90.97 dB
With 1 pF the response is flat, but the band ends at 3.27 MHz.
- Frequency 15.9 MHzno CF = 105.5 dB · CF = 0.25 pF = 90.97 dB · CF = 1 pF = 79.66 dB
With 0.25 pF it is flat to 15.9 MHz.
- Frequency 17.78 MHzno CF = 108.7 dB · CF = 0.25 pF = 89.9 dB · CF = 1 pF = 78.62 dB
With no CF the response peaks by 14.73 dB at 17.78 MHz.
Read the plot as a table
| Frequency | no CF | CF = 0.25 pF | CF = 1 pF |
|---|---|---|---|
| 10 kHz | 93.97 dB | 93.97 dB | 93.97 dB |
| 100 kHz | 93.97 dB | 93.97 dB | 93.97 dB |
| 1 MHz | 94 dB | 93.97 dB | 93.59 dB |
| 3.269 MHz | 94.26 dB | 93.96 dB | 90.97 dB |
| 10 MHz | 97.15 dB | 93.33 dB | 83.69 dB |
| 15.9 MHz | 105.5 dB | 90.97 dB | 79.66 dB |
| 17.78 MHz | 108.7 dB | 89.9 dB | 78.62 dB |
| 100 MHz | 64.31 dB | 62.09 dB | 57.1 dB |
| 1 GHz | 24.04 dB | 22.1 dB | 18.01 dB |
The op-amp is a behavioural macro with one pole.
The resistor set the gain correctly. The loop around it set everything else.
Shunt-Shunt Feedback Sets the Gain
A photodiode delivers a current. A transimpedance amplifier (TIA) turns that current into a voltage. Its gain is in ohms.
- The loop holds the input still
The amplifier moves its output until RF carries all of ipd. So vout = −ipd·RF.
ipd · ipd
- CD steals current
The input node must move a little. At high frequency CD takes part of the current.
iC
- CF shapes the loop
CF passes fast signals around RF. It trades bandwidth for a flat response.
Read the circuit as text
- i_pd current source: + →
0, − →in - C_D capacitor (1 pF): pin 1 →
in, pin 2 →0 - A op-amp: −in →
in, +in →0, out →n1 - R_F resistor (50 kΩ): pin 1 →
n1, pin 2 →in - C_F capacitor (0.25 pF): pin 1 →
n1, pin 2 →in - out port: terminal →
n1
The network senses the output voltage and returns a current to the input node.
E10.1 named this loop shunt-shunt. A shunt connection lowers the resistance at both ports by 1 + T.
The lab drives 1 A of AC into each input. Then v(out) reads the transimpedance and v(in) reads the input resistance.
| Amplifier | Gain A | Transimpedance | Input resistance |
|---|---|---|---|
| Op-amp macro | 1000 | 49.95 kΩ | 49.95 Ω |
| One inverter | 66.08 | 48.37 kΩ | 1.626 kΩ |
Both use RF = 50 kΩ. Read at 1 kHz.
With a gain of 1000 the transimpedance is 0.1 % under RF. The input looks like 50 Ω, which is RF/(1 + A).
The inverter has a gain of only 66, and RF also loads its output. By hand, (gm·RF − 1)/(gm + gds) gives zt_b_hand = 4.837981e+04.
The Photodiode Capacitance and the Loop
Without an amplifier, RF and CD alone would roll off at 1/(2π·RF·CD) = 3.18 MHz. The loop lowers the input resistance, so the band is wider.
But RF and CD also sit inside the loop. They add a pole to the loop gain at 3.18 MHz. The op-amp adds its own pole.
- Source
- I/O
- Device
- Measure
- Data
- The gain
The loop makes RF carry all of ipd. So RF sets the gain.
- Two poles
Once around the loop the signal meets two poles. One is RF with CD. The other is in the amplifier.
- One zero
CF passes fast signals around RF. That adds a zero to the loop gain.
Read the diagram as text
- Photodiode: current i_pd
- Input node: CD = 1 pF to ground
- Amplifier: gain A, one pole
- Output: a voltage
- RF: 50 kΩ, sets the gain
- CF: adds a zero
- Photodiode → Input node: current
- Input node → Amplifier: small voltage
- Amplifier → Output
- Output → RF: senses voltage
- RF → Input node: returns current
- Output → CF
- CF → Input node: fast current
The same loop serves both amplifiers of this lesson.
Two poles in a loop make a resonance. For an op-amp with gain-bandwidth GBW, it sits at f0 = √(GBW·fRC), where fRC is the 3.18 MHz above.
Here GBW is 100 MHz, so f0 = 17.8 MHz. The run peaks at fpeak_a_nocf = 1.778279e+07, by peak_a_nocf = 1.473261e+01 dB.
A Feedback Capacitor for a Flat Response
A capacitor CF across RF adds a zero to the loop gain at 1/(2π·RF·CF). The zero returns phase near f0, so the peak shrinks.
For a maximally flat response, the rule is CF ≈ √(2·CD/(2π·GBW·RF)). Here that is 0.25 pF.
This rule aims at Q ≈ 0.707, a Butterworth response. Many app notes drop the 2 and aim at about 45° of phase margin instead.
- no CF
- CF = 0.25 pF
- CF = 1 pF
- Time 36.6 nsno CF = -86.73 mV · CF = 0.25 pF = -45 mV · CF = 1 pF = -19.01 mV
With 0.25 pF the output reaches 90 % of its step 26.6 ns after the edge.
- Time 38.15 nsno CF = -87.31 mV · CF = 0.25 pF = -46.47 mV · CF = 1 pF = -19.98 mV
With no CF the output swings to −87.3 mV, 75 % past its final value. It then rings every 56.3 ns.
- Time 54.45 nsno CF = -45.04 mV · CF = 0.25 pF = -52.07 mV · CF = 1 pF = -28.56 mV
With 0.25 pF the lowest point is −52.07 mV: 4.2 % of overshoot.
- Time 124.8 nsno CF = -34.79 mV · CF = 0.25 pF = -49.94 mV · CF = 1 pF = -44.96 mV
With 1 pF there is no overshoot, but 90 % takes 114.8 ns.
Read the plot as a table
| Time | no CF | CF = 0.25 pF | CF = 1 pF |
|---|---|---|---|
| 0 s | 0 V | 0 V | 0 V |
| 36.6 ns | -86.73 mV | -45 mV | -19.01 mV |
| 38.15 ns | -87.31 mV | -46.47 mV | -19.98 mV |
| 50 ns | -61.08 mV | -51.81 mV | -26.5 mV |
| 54.45 ns | -45.04 mV | -52.07 mV | -28.56 mV |
| 100 ns | -67.06 mV | -49.86 mV | -41.61 mV |
| 124.8 ns | -34.79 mV | -49.94 mV | -44.96 mV |
| 150 ns | -61.62 mV | -49.95 mV | -46.99 mV |
| 200 ns | -54.52 mV | -49.95 mV | -48.9 mV |
| 250 ns | -49.96 mV | -49.95 mV | -49.58 mV |
| 300 ns | -48.33 mV | -49.95 mV | -49.82 mV |
| 350 ns | -48.47 mV | -49.95 mV | -49.9 mV |
| 400 ns | -49.17 mV | -49.95 mV | -49.93 mV |
A flat response in frequency is a clean step in time.
| C_F | Peaking | −3 dB frequency |
|---|---|---|
| none | 14.73 dB | 27.56 MHz |
| 0.25 pF | 0 dB | 15.90 MHz |
| 1 pF | 0 dB | 3.269 MHz |
The lab prints these rows for the op-amp TIA.
Too much CF wastes the amplifier. With 1 pF the band is just 1/(2π·RF·CF), about 3.2 MHz.
A TIA from One Inverter
- The amplifier
Both transistors amplify. The gain is (gmn + gmp)/(gdsn + gdsp) = 66.
- The loop
RF closes the same shunt-shunt loop. The second pole now comes from CL at the output.
ipd
Read the circuit as text
- i_pd current source: + →
0, − →n1 - C_D capacitor (1 pF): pin 1 →
n1, pin 2 →0 - MP PMOS transistor (16/0.3): gate →
n1, source →VDD, drain →n2 - MN NMOS transistor (8/0.3): gate →
n1, drain →n2, source →0 - R_F resistor (50 kΩ): pin 1 →
n2, pin 2 →n1 - C_F capacitor (80 fF): pin 1 →
n2, pin 2 →n1 - C_L capacitor (0.3 pF): pin 1 →
n2, pin 2 →0 - out port: terminal →
n2
It draws 43.1 µA from 1.8 V.
| C_F | Peaking | −3 dB frequency |
|---|---|---|
| none | 5.39 dB | 62.14 MHz |
| 80 fF | 0.0004 dB | 36.85 MHz |
The inverter TIA with its 0.3 pF load.
One inverter reaches 36.8 MHz, more than twice the band of the 100 MHz op-amp. Its 1 µA step overshoots by 4.4 %.
The price is accuracy. The gain is 3.3 % under RF, and it moves with process and supply.
The Noise of the Feedback Resistor
A resistor R makes a thermal noise current of √(4kT/R) (E7.1). The noise of RF flows into the input node, like the signal.
So RF alone gives an input-referred noise current of √(4kT/RF). For 50 kΩ that is in_rf_hand = 5.757789e-13, or 0.576 pA/√Hz.
noise with the photodiode current source as the input.- inverter TIA
- op-amp TIA
- Frequency 100 kHzinverter TIA = 1.023 pA/rtHz · op-amp TIA = 575.8 fA/rtHz
At 100 kHz the inverter TIA has 1.023 pA/√Hz. Flicker noise of the transistors adds to RF.
- Frequency 1 MHzinverter TIA = 682.8 fA/rtHz · op-amp TIA = 575.8 fA/rtHz
At 1 MHz it has 0.683 pA/√Hz. The op-amp TIA has 0.576 pA/√Hz: RF alone.
- Frequency 30 MHzinverter TIA = 1.435 pA/rtHz · op-amp TIA = 576.6 fA/rtHz
At 30 MHz the inverter TIA is back up at 1.435 pA/√Hz.
Read the plot as a table
| Frequency | inverter TIA | op-amp TIA |
|---|---|---|
| 1 kHz | 5.85 pA/rtHz | 575.8 fA/rtHz |
| 10 kHz | 2.278 pA/rtHz | 575.8 fA/rtHz |
| 100 kHz | 1.023 pA/rtHz | 575.8 fA/rtHz |
| 1 MHz | 682.8 fA/rtHz | 575.8 fA/rtHz |
| 10 MHz | 814.9 fA/rtHz | 575.9 fA/rtHz |
| 30 MHz | 1.435 pA/rtHz | 576.6 fA/rtHz |
| 100 MHz | 3.832 pA/rtHz | 584.6 fA/rtHz |
| 1 GHz | 30.66 pA/rtHz | 1.164 pA/rtHz |
The macro op-amp adds almost no noise, so its curve is the noise of RF. rtHz is √Hz.
The rise at high frequency has a simple cause. The amplifier's own noise voltage appears across CD, and a capacitor passes more current at higher frequency.
Asked in interviews
Why does a TIA peak or ring when the photodiode capacitance is large?
RF and the input capacitance add a pole inside the loop. With the amplifier's own pole, the loop has two poles and little phase margin. A small capacitor across RF adds a zero that restores the margin.
How do you choose RF?
Use the largest value that still meets the bandwidth. The signal grows with RF, but its noise voltage grows only with √RF. The resonance falls as 1/√RF, so each doubling of RF costs about 30 % of the band.
What is the input resistance of a TIA?
About RF/(1 + A), because the feedback is in shunt with the input. The lab measures 50 Ω with a gain of 1000 and 1.63 kΩ with one inverter.
Try It
Each photodiode is a current source with 1 pF beside it. Its 1 A of AC makes v(out) read the transimpedance in ohms and v(in) the input resistance.
Part 2 shrinks both feedback capacitors to 1 aF with alter and reads the peaking. Part 3 tries 1 pF on the op-amp TIA.
Part 5 runs noise with the current source as the input, so inoise_spectrum is already a current density. An ac run on the same frequencies carries the vectors, because meas cannot read a noise plot.
Predict before each change. Set CDA to 4p: where does the peak move, and what CF flattens it? Set CLB to 1p: does 80 fF still give a flat response?
Run the netlist to draw its plots here.
zt_a = 4.995005e+04 rin_a = 4.995255e+01 f3_a = 1.589842e+07 zt_b = 4.837355e+04 rin_b = 1.626451e+03 f3_b = 3.684534e+07 a_inv = 6.608436e+01 zt_b_hand = 4.837981e+04 f3_a_nocf = 2.756014e+07 f3_b_nocf = 6.214395e+07 f3_a_1p = 3.268826e+06 step_a = -4.995324e-02 step_b = -4.837039e-02 in_rf_hand = 5.757789e-13 in_a_1meg = 5.757855e-13 in_b_100k = 1.023171e-12 in_b_1meg = 6.828128e-13 in_b_30meg = 1.434743e-12Check Yourself
Common Mistakes
Wrong Leave CF out because the op-amp is stable as a unity-gain buffer.
RF and the photodiode capacitance add a pole inside the loop. The lab's TIA peaks by 14.7 dB without CF.
Fix Count the input capacitance as part of the loop and add a CF near √(2·CD/(2π·GBW·RF)).
Wrong Make CF several times larger than needed, to be safe.
The band then falls to 1/(2π·RF·CF). With 1 pF the band falls from 15.9 MHz to 3.27 MHz.
Fix Use the smallest CF that meets the peaking limit, and check it with a step.
Wrong Lower RF to lower the noise.
The noise current of RF is √(4kT/RF), so it rises as RF falls.
Fix Choose the largest RF that the bandwidth allows.
Wrong Read onoise_spectrum and call it the noise of the TIA.
That is a voltage at the output. It cannot be compared with a photodiode current.
Fix Name the photodiode current source as the input of noise and read inoise_spectrum, in A/√Hz.
noise v(outa) IA dec 100 1k 1g
let n = noise1.onoise_spectrumnoise v(outa) IA dec 100 1k 1g
let n = noise1.inoise_spectrumRemember
- A TIA is a shunt-shunt loop: vout = −i·RF, and the input resistance is about RF/(1 + A).
- RF and the photodiode capacitance add a pole inside the loop. The resonance sits near √(GBW·fRC).
- CF adds a zero that removes the peak. Too much CF cuts the band to 1/(2π·RF·CF).
- RF gives an input-referred noise current of √(4kT/RF). The amplifier adds flicker noise below and a rise above.