Making small signals big
A microphone produces a few millivolts. A guitar pickup, an electrocardiogram electrode, a radio antenna and a strain gauge produce signals of the same size or smaller. The converter that will turn the signal into numbers wants a volt or more, and a loudspeaker wants watts. In between sits an amplifier, which makes a copy of the signal that is larger but otherwise as nearly identical as possible.
The previous lesson used transistors as switches, slammed fully on or fully off. An amplifier uses them gently, in the middle of their range, where a small change at the base or gate makes a proportionally larger change in the collector or drain current. The difficulty is that the devices are nonlinear, exponential (BJT) or square-law (MOSFET). This lesson's central idea is that any smooth curve looks straight if you look at a small enough piece of it. Choose a quiet operating point, keep the signal small, and the transistor becomes a linear circuit element whose gain you can calculate with the tools of the Circuits level.
Biasing: choosing the operating point
Biasing sets the DC voltages and currents around which the signal will swing: the operating point, or quiescent point (Q-point). For a BJT amplifier that means a collector current and a collector–emitter voltage that keep it in the forward-active region with room for the signal to swing both ways: the collector must not fall into saturation on the negative swings, nor rise to the supply on the positive ones.
The naive way to bias a BJT is a single resistor from the supply to the base, setting and hence . It fails, because varies by a factor of two or three between parts and with temperature. The standard cure is the four-resistor bias: a divider fixes the base voltage, and an emitter resistor turns that voltage into a current: If the transistor's current rises, the emitter voltage rises, falls, and the current is pulled back: negative feedback, the subject of Feedback, stability and oscillators. Replacing the divider by its Thévenin equivalent ( behind , from Network theorems) gives the exact relation which depends on only through the small term .
import math
VT, IS, VA = 0.025852, 1e-14, 100.0 # thermal voltage, saturation current, Early voltage
VCC, R1, R2, RC, RE = 12.0, 47e3, 10e3, 4.7e3, 1e3
def bias(beta):
"""Solve the four-resistor bias circuit: Thevenin base source, exponential V_BE."""
vth, rth = VCC * R2 / (R1 + R2), R1 * R2 / (R1 + R2)
ic = 1e-3
for _ in range(50): # fixed-point iteration on V_BE (converges fast)
vbe = VT * math.log(ic / IS)
ie = (vth - vbe) / (RE + rth / (beta + 1))
ic = ie * beta / (beta + 1)
return ic, vbe, VCC - ic * RC - ie * RE
print("four-resistor bias (R1 47k, R2 10k, RC 4.7k, RE 1k, VCC 12 V)")
for beta in (100, 200, 400):
ic, vbe, vce = bias(beta)
print(f" beta {beta:3d}: I_C = {ic * 1e3:.3f} mA, V_BE = {vbe:.3f} V, V_CE = {vce:.2f} V")
print("one base resistor from VCC instead, sized for 1.38 mA at beta = 200")
RB = (VCC - 0.66) / (1.378e-3 / 200)
for beta in (100, 200, 400):
ic = beta * (VCC - 0.66) / RB
vce = VCC - ic * RC
if vce > 0.2:
print(f" beta {beta:3d}: I_C = {ic * 1e3:.3f} mA, V_CE = {vce:.2f} V")
else:
print(f" beta {beta:3d}: would need V_CE = {vce:.2f} V, impossible: the transistor saturates")
ic, vbe, vce = bias(200)
gm = ic / VT
rpi = 200 / gm
ro = VA / ic
print(f"\nsmall-signal parameters at beta = 200: gm = {gm * 1e3:.1f} mS, "
f"r_pi = {rpi / 1e3:.2f} kohm, r_o = {ro / 1e3:.1f} kohm, gm*r_o = {gm * ro:.0f}")
RL = 10e3
rload = 1 / (1 / RC + 1 / ro + 1 / RL)
print(f"midband gain with RE bypassed, 10 kohm load: -gm (RC || r_o || RL) = {-gm * rload:.1f}")
print(f"with RE left unbypassed: about -(RC || RL) / (RE + 1/gm) = {-(1 / (1 / RC + 1 / RL)) / (RE + 1 / gm):.2f}")With four-resistor bias, a fourfold spread in , from 100 to 400, moves the collector current only from 1.321 to 1.409 mA, under 7%. With a single base resistor, the same spread takes the current from 0.689 mA to a value the circuit cannot reach: the transistor saturates and the amplifier stops working. The rest of the output looks ahead to the small-signal parameters, which the next section defines.
Small-signal models
Write each quantity as its DC operating value plus a small signal: , . Expanding the exponential in a Taylor series around the operating point, If is a few millivolts, much less than mV, the squared term is small and : the signal current is proportional to the signal voltage.
Definition— Transconductance and the small-signal parameters
The transconductance at the operating point is the transistor's gain, in amperes per volt (siemens). With it come the input resistance seen at the base and the output resistance seen at the collector or drain:
| BJT | MOSFET (square law) | |
|---|---|---|
| input | infinite (insulated gate) | |
| ( = Early voltage) | ||
| intrinsic gain | , often 1000–5000 | , often 10–100 |
The striking fact about a BJT is that depends only on the current: every BJT at 1 mA has mS, whatever its size or maker. A MOSFET's depends on its size and overdrive too, and for the same current it is smaller: at 1 mA with a typical overdrive of 0.2 V, mS.
The hybrid-π model puts these together into a linear circuit that replaces the transistor in signal calculations, with two capacitances added for high frequencies: between base and emitter (the depletion capacitance of the junction plus the charge stored in the base), and across the reverse-biased collector–base junction. For a MOSFET, is infinite, plays the part of and the part of .
Common-emitter and common-source gain
In the amplifier of the figure, the signal enters at the base and leaves at the collector, with the emitter grounded for signals by : a common-emitter (CE) stage. Replace the transistor by its hybrid-π model and every capacitor by a short at midband frequencies. The input voltage appears directly across , so the collector current is , and it flows through everything connected to the collector in parallel, : The minus sign means the stage inverts: when the base rises, more current flows through and the collector falls. For our circuit, mS and kΩ, so , as the program printed. The common-source (CS) stage of a MOSFET gives in exactly the same way.
Example— Trading gain for predictability: emitter degeneration
Leave unbypassed and the signal sees it too. The input now divides between the junction and , and the gain becomes With and kΩ, the gain drops to about . That looks like a loss, but notice what has gone from the formula: almost all dependence on , which changes with temperature, with the operating point, and over the signal's own swing, which causes distortion. The gain is now set by a ratio of resistors. A common compromise bypasses part of and leaves a few tens of ohms unbypassed, for a gain of perhaps 30 to 50 that is both useful and predictable. This is negative feedback again: giving up gain to buy accuracy and linearity.
The best gain a single transistor can give is with a load much larger than its own : the intrinsic gain . A BJT's is , about 3900 in our example. A modern short-channel MOSFET may manage only 10 to 30, which is why integrated amplifiers replace load resistors by current-source loads (transistors whose output resistance is high) and stack transistors in cascodes to multiply output resistance.
Followers: gain of one, and why that is useful
Take the output from the emitter instead of the collector, with the collector tied to the supply: a common-collector stage, better called an emitter follower. The emitter follows the base, one lower, so the voltage gain is just under one: where is everything connected to the emitter. The follower's value is in its resistances: it presents a high resistance to the source and a low one to the load. It is a buffer, an impedance converter.
Example— Driving a 100 Ω load from a 10 kΩ source
Connect a 100 Ω load directly to a source with 10 kΩ of internal resistance, and the load receives only of the signal. Put an emitter follower biased at 5 mA between them, with : its output resistance is , so the load now receives of the signal. The follower has no voltage gain, yet it multiplied the signal at the load by about 65.
The MOSFET equivalent, the source follower, has infinite input resistance and an output resistance of about . Its gain is a little further below one, because the body effect raises the threshold as the source rises. Followers appear at the outputs of op-amps, as level shifters, and as the buffers of voltage references.
The differential pair
Most amplifiers inside chips do not amplify a voltage relative to ground. They amplify the difference between two inputs, and ignore what the two have in common. The circuit that does this is the differential pair: two matched transistors whose emitters (or sources) are joined and fed by a single tail current .
Because the tail current is fixed, whatever one transistor gains the other loses. With bipolar transistors the exponential law makes the split exact: the ratio of the collector currents is , where is the difference between the base voltages, and the differential output is For small , and the gain is with of one transistor at half the tail current. A voltage applied equally to both bases, a common-mode signal such as hum picked up on both wires, moves both emitters together and leaves the split unchanged. With an imperfect tail source of output resistance , a little common-mode signal leaks through, with gain about at each output. The ratio of differential to common-mode gain is the common-mode rejection ratio (CMRR).
import math
VT = 0.025852
I_TAIL, RC = 1e-3, 10e3 # tail current and collector resistors
R_TAIL = 200e3 # output resistance of a real (imperfect) tail current source
def collector_currents(vd):
"""Large-signal BJT pair: the tail current divides according to exp(V_BE / V_T)."""
ic1 = I_TAIL / (1 + math.exp(-vd / VT))
return ic1, I_TAIL - ic1
print(" v_d (mV) I_C1 (mA) I_C2 (mA) v_out (V) linear model (V)")
gm = I_TAIL / 2 / VT # each transistor carries half the tail current
for vd_mv in (0, 5, 10, 20, 40, 60, 100, 200):
ic1, ic2 = collector_currents(vd_mv / 1e3)
vout = (ic1 - ic2) * RC # differential output, (V_C2 - V_C1)
print(f"{vd_mv:8d} {ic1 * 1e3:9.4f} {ic2 * 1e3:9.4f} {vout:9.4f} {gm * RC * vd_mv / 1e3:12.4f}")
print(f"\ndifferential gain gm*RC = {gm * RC:.1f}; exact output = I_tail*RC*tanh(v_d / 2V_T)")
acm = RC / (2 * R_TAIL) # each side, for a common-mode input
print(f"common-mode gain (single-ended) about RC / (2 R_tail) = {acm:.3f}")
adm = gm * RC / 2 # single-ended differential gain
print(f"CMRR = A_dm / A_cm = {adm / acm:.0f}, or {20 * math.log10(adm / acm):.1f} dB")The pair is linear only for up to about 10 mV (within about 1.2% of the straight line there); at 40 mV it delivers 6.5 V instead of the linear 7.7 V, and beyond about 100 mV it simply steers the whole tail current to one side. That gentle, symmetric saturation is very useful: a differential pair overdriven by a large input limits cleanly instead of breaking down. The CMRR of 3868, about 72 dB, comes from a tail source of only 200 kΩ; integrated designs use better current sources and reach 80 to 120 dB. Every operational amplifier begins with a differential pair, which is how the next lesson starts.
Frequency response and the Miller effect
An amplifier's gain is flat only over a band of frequencies. At low frequencies the coupling and bypass capacitors stop being short circuits; at high frequencies the transistor's own capacitances and start to conduct. Each capacitor contributes a pole, as in Frequency response and filters, and the two corners where the gain falls by 3 dB bound the bandwidth.
The low corner here is set by the emitter bypass capacitor, even though it is the largest. It must short out against the resistance looking into the emitter, which is tiny, about . So 100 µF gives a corner near Hz, while the 10 µF coupling capacitors, which see kilohms, contribute corners of only a few hertz.
The high corner is dominated by the small capacitance , because of the Miller effect, first described for vacuum tubes by John Milton Miller around 1920. connects the input to an output that swings times as far in the opposite direction. The voltage across it is therefore times the input voltage, so it draws times as much current as a capacitor to ground would: from the input it looks like where . With pF and a gain of 163, that is about 657 pF, twenty times the of 31 pF. The input pole is then at , with .
The program below checks that estimate against an exact AC analysis of the whole small-signal circuit: it writes the nodal equations at each frequency, with complex admittances for every capacitor, and solves them with NumPy, the method of Nodal and mesh analysis and exactly what SPICE's .ac does.
import math
import numpy as np
# Small-signal values from the bias point (beta = 200, I_C = 1.378 mA)
gm, rpi, ro = 1.378e-3 / 0.025852, 200 / (1.378e-3 / 0.025852), 100 / 1.378e-3
RS, RTH, RC, RE, RL = 1e3, 47e3 * 10e3 / 57e3, 4.7e3, 1e3, 10e3
CIN, COUT, CE = 10e-6, 10e-6, 100e-6
CPI, CMU = 10e-12 + 0.4e-9 * gm, 4e-12 # junction + diffusion capacitance; collector junction
def gain(f):
"""Solve the hybrid-pi circuit's nodal equations at frequency f; return v_out / v_source."""
s = 2j * math.pi * f
n = 5 # nodes: 0 source side of C_in, 1 base, 2 emitter, 3 collector, 4 output
Y = np.zeros((n, n), complex)
I = np.zeros(n, complex)
def admit(a, b, y): # an admittance y between nodes a and b (None = ground)
for p, q in ((a, b), (b, a)):
if p is not None:
Y[p, p] += y
if q is not None:
Y[p, q] -= y
admit(0, None, 1 / RS); I[0] = 1 / RS # 1 V source behind R_S, as its Norton equivalent
admit(0, 1, s * CIN)
admit(1, None, 1 / RTH)
admit(1, 2, 1 / rpi + s * CPI)
admit(1, 3, s * CMU)
admit(3, 2, 1 / ro)
admit(3, None, 1 / RC)
admit(2, None, 1 / RE + s * CE)
admit(3, 4, s * COUT)
admit(4, None, 1 / RL)
Y[3, 1] += gm; Y[3, 2] -= gm # g_m v_be flows from collector to emitter
Y[2, 1] -= gm; Y[2, 2] += gm
return np.linalg.solve(Y, I)[4]
freqs = [10, 30, 70, 200, 1e3, 1e4, 1e5, 3e5, 1e6, 3e6]
print(" f (Hz) |gain| gain (dB) phase (deg)")
for f in freqs:
g = gain(f)
print(f"{f:9.0f} {abs(g):8.1f} {20 * math.log10(abs(g)):8.1f} {math.degrees(np.angle(g)):9.1f}")
mid = abs(gain(1e4))
def corner(lo, hi): # bisection on a log scale for |gain| = mid / sqrt 2
for _ in range(60):
m = math.sqrt(lo * hi)
if (abs(gain(m)) > mid / math.sqrt(2)) == (abs(gain(lo)) > mid / math.sqrt(2)):
lo = m
else:
hi = m
return m
print(f"\nmidband gain {mid:.1f}; -3 dB at {corner(1, 1e4):.1f} Hz and {corner(1e4, 1e8) / 1e3:.0f} kHz")
rload = 1 / (1 / RC + 1 / ro + 1 / RL)
cmiller = CMU * (1 + gm * rload)
rsig = 1 / (1 / RS + 1 / RTH + 1 / rpi)
print(f"Miller estimate: C_mu x (1 + gm R'L) = {cmiller * 1e12:.0f} pF, plus C_pi {CPI * 1e12:.1f} pF, "
f"seen through {rsig:.0f} ohm -> f_H = {1 / (2 * math.pi * rsig * (cmiller + CPI)) / 1e3:.0f} kHz")From the source the midband gain is 117.6 (the 163 of the transistor stage, reduced because the 1 kΩ source resistance and the 2.6 kΩ input resistance form a divider). The band runs from 70.3 Hz to 314 kHz, and the one-line Miller estimate, 321 kHz, is within 3%. Notice the phase: −180° in midband (the inversion), moving towards −90° at each end as each pole adds or removes up to 90°.
The check in SPICE
Now the whole circuit, bias and all, with SPICE's own BJT model in place of our hand-built small-signal one. The .model line gives a plain textbook npn with plausible parameters, not a manufacturer's part: fA, , an Early voltage of 100 V, junction capacitances of 10 and 4 pF held constant (MJE=0, MJC=0) to match the hand calculation, and a forward transit time of 0.4 ns for the charge stored in the base.
* Common-emitter amplifier: bias point and frequency response
VCC vcc 0 DC 12
VS sig 0 DC 0 AC 1
RS sig in 1k
CIN in b 10u
R1 vcc b 47k
R2 b 0 10k
Q1 c b e QNPN
RC vcc c 4.7k
RE e 0 1k
CE e 0 100u
COUT c out 10u
RL out 0 10k
* A textbook small-signal npn with plausible parameters (not a vendor model).
* MJE=MJC=0 keeps the junction capacitances constant, as in the hand calculation.
.model QNPN NPN(IS=1e-14 BF=200 VAF=100 CJE=10p MJE=0 CJC=4p MJC=0 TF=0.4n)
.options TEMP=26.85 TNOM=26.85
.control
op
print @q1[ic] v(c,e) v(b,e) @q1[gm] @q1[cmu] @q1[cpi]
ac dec 50 1 100Meg
let gain = abs(v(out))
meas ac gmid MAX gain from=1k to=100k
let g3 = gmid / sqrt(2)
meas ac flow WHEN gain=g3 RISE=1
meas ac fhigh WHEN gain=g3 FALL=1
echo "midband gain $&gmid, -3 dB at $&flow Hz and $&fhigh Hz"
.endc
.end| Quantity | Hand and Python | SPICE |
|---|---|---|
| 1.378 mA | 1.382 mA | |
| 4.14 V | 4.12 V | |
| 53.3 mS | 53.4 mS | |
| 31.3 pF | 31.4 pF | |
| midband gain from the source | 117.6 | 118.8 |
| low −3 dB frequency | 70.3 Hz | 71.0 Hz |
| high −3 dB frequency | 314 kHz (Miller: 321 kHz) | 311 kHz |
The agreement is within about 1%. The small differences come from things the hand analysis ignored: SPICE includes the Early effect in the bias point (raising slightly) and its Gummel–Poon model makes itself depend a little on . This is the professional rhythm: estimate by hand, so you know what answer to expect and which component matters, then simulate to get the details right.
The Miller effect has well-known cures. A cascode stacks a common-base transistor on top of the CE stage, so the first transistor's collector sees only the low input resistance of the second and has almost no voltage gain to multiply . Followers have no Miller multiplication at all. And the transistor's own speed limit is its transition frequency , the frequency at which its current gain falls to one: 240 MHz for our model transistor, and hundreds of gigahertz for the silicon–germanium transistors in radio chips.
Amplifiers in professional practice
Discrete single-transistor amplifiers like this one are now rare outside teaching, audio enthusiasm and radio-frequency front ends; most products use integrated operational amplifiers and dedicated amplifier chips. But every one of those chips is built from the stages in this lesson. An analog integrated-circuit designer sizes transistors for a chosen (transconductance per unit of current, a measure of how efficiently a transistor turns bias current into gain), biases them with current mirrors, uses differential pairs everywhere for their rejection of supply and substrate noise, adds cascodes for gain, and simulates across process, voltage and temperature corners with the foundry's BSIM models before committing to silicon.
You can now do this yourself with open tools: ngspice, the open-source process design kits SkyWater SKY130 (from SkyWater and Google, United States), GlobalFoundries GF180MCU, and IHP SG13G2 (a silicon–germanium process from Germany's IHP), and community programmes that fabricate small student designs. How chips are designed describes the full flow.
Common misconceptions
Misconception
"More gain is always better." Gain beyond what the signal needs only amplifies noise and brings the stage closer to clipping, and high-gain stages are the most sensitive to temperature, part variation and the Miller effect. Engineers usually choose modest, accurate gain, and get it with feedback.
Misconception
" is a property of the transistor type, like β on the datasheet." A BJT's is set by its bias current, , and a MOSFET's by its current and overdrive. Change the bias and you change the gain.
Misconception
"The small-signal model works for any signal size." It is a tangent-line approximation, accurate only while stays well below (or well below the overdrive). Larger signals distort, as the differential pair's tanh curve showed.
Misconception
"A follower is useless because its gain is one." Its power gain and current gain are large. It lets a weak, high-resistance source drive a heavy load, which in the example multiplied the signal at the load by about 65.
Exercises
1. Transistor amplifiers
Warm-upHumanA BJT is biased at mA at room temperature ( mV). What is its transconductance, in millisiemens, to the nearest whole number?
2. Transistor amplifiers
Warm-upHumanA BJT with is biased at mA. Taking mV, what is in ohms?
3. Transistor amplifiers
CoreHumanA common-emitter stage is biased at mA (take mV), with kΩ, an unbypassed emitter resistor , no external load, and large enough to ignore. What is the magnitude of its voltage gain, to the nearest whole number?
4. Transistor amplifiers
CoreHumanA common-source stage has a gate–drain capacitance pF and a midband voltage gain of . What capacitance, in picofarads, does appear to add at the input?
5. Transistor amplifiers
CoreHumanA sensor with 50 kΩ of internal resistance must drive a 200 Ω cable termination. The signal is already large enough; you only need it to arrive undiminished. Which single-transistor stage fits best?
6. Transistor amplifiers
CoreHumanA BJT differential pair has a tail current source and equal collector resistors. Which statements are true? Select all that apply.
7. Transistor amplifiers
CoreComputationalWrite ce_bias(vcc, r1, r2, rc, re, beta, vbe=0.7) for the four-resistor bias circuit of the lesson (divider from the supply to the base and from the base to ground, in the collector, in the emitter). Treat as fixed. Return (ic, vce, region) with region equal to "active" when V and "saturated" otherwise (the values you return in the saturated case are not checked).
Use the Thévenin equivalent of the divider and include the base current exactly: .
Further reading and resources
- ngspice documentationThe manual describes the Gummel–Poon BJT model and the operating-point and AC analyses used here.
- SkyWater SKY130 open-source process design kitA real 130 nm CMOS process, free to design with, including its transistor models.
- IHP SG13G2 open-source process design kitA 130 nm silicon–germanium BiCMOS process from IHP in Germany, open for anyone to design analog and radio circuits.
- BSIM Group, University of California, BerkeleyThe compact MOSFET models behind professional analog simulation.
- MIT OpenCourseWare: Circuits and Electronics (6.002)Lecture notes and problems that cover small-signal models and amplifiers.