Skip to main content

Transconductance and output resistance

You’ll often hear that a transistor amplifies. This page puts a number on how much it can amplify, and shows that what sets the number is the product \(g_m r_o\).

So far you’ve measured currents, and each one answered the question “what does this device do at this bias?”. An amplifier cares about something else: how much the answer changes when you nudge the input. That’s a slope rather than a value, and a transistor has two slopes that matter, one for each terminal you can move.

Symbol Name It is the slope of In words
\(g_m\) transconductance \(I_D\) vs \(V_{GS}\) how hard the gate pushes current through the device
\(g_{ds}\) output conductance \(I_D\) vs \(V_{DS}\) how much the drain voltage leaks through to the current
\(r_o = 1/g_{ds}\) output resistance — the same slope, written as its reciprocal

\(g_m\) is measured in siemens because it’s amps per volt, and it’s called trans-conductance because the volts are at one terminal and the amps at another. That’s the reason a transistor can have gain at all. A two-terminal device’s conductance can only turn a voltage across it into the current through itself, but a transistor lets a voltage on the gate control the current through the drain.

Measuring \(g_m\)

\(g_m\) is the slope of the \(I_D\) vs \(V_{GS}\) curve you swept in Lab 03. When you differentiated that sweep in Lab 03, Part 5, you got:

== g_m = dI_D/dV_GS   (V_DS = 1.8 V)
   V_GS = 0.90 V   I_D =    63.876 uA   g_m =  376.932 uS   g_m/I_D =   5.90 /V
   V_GS = 1.20 V   I_D =   217.505 uA   g_m =  634.008 uS   g_m/I_D =   2.91 /V
   V_GS = 1.50 V   I_D =   436.080 uA   g_m =  809.693 uS   g_m/I_D =   1.86 /V
   V_GS = 1.80 V   I_D =   696.275 uA   g_m =  914.650 uS   g_m/I_D =   1.31 /V

and the same lab printed it next to the model’s own value:

   g_m at V_GS = 1.8 V                914.6500   915.3117  uS   -0.07 %

Your 361-point sweep gives 914.650 µS, and BSIM’s analytic derivative gives 915.3117 µS, so the two are only 0.07 % apart. Compare that with the threshold voltage, where two reasonable methods disagreed by 86 mV. The difference is that \(g_m\) is simply the slope of a measured curve, so there’s no choice of definition to argue about.

The square-law estimate

Differentiating the square law gives a tidy expression for the slope:

\[I_D = \tfrac{1}{2}\mu_n C_{ox}\tfrac{W}{L}V_{ov}^2 \quad\Longrightarrow\quad g_m = \frac{2 I_D}{V_{ov}}\]

At \(V_{GS} = 1.8\) V, with \(V_{TH} = 0.6016\) V from Where saturation starts, the overdrive is 1.1984 V and \(I_D\) is 696.275 µA:

\[g_m = \frac{2 \times 696.275\ \mu\text{A}}{1.1984\ \text{V}} = 1162.0\ \mu\text{S}\]

The model gives 915.3 µS, so the formula comes out 27 % high.

Lower down, at \(V_{GS} = 0.9\) V, the overdrive is only 0.2984 V:

\[g_m = \frac{2 \times 63.876\ \mu\text{A}}{0.2984\ \text{V}} = 428.1\ \mu\text{S}\]

The measured value there is 376.9 µS, so at this lower bias the formula is only 13.6 % high.

This is the same way the square law failed at the saturation knee, and it has the same cause, velocity saturation. As you push the overdrive up, the current stops growing quadratically, so its derivative stops being \(2I_D/V_{ov}\). In the fully velocity-saturated limit \(g_m = I_D/V_{ov}\), which at 1.8 V would predict 581 µS, 36 % low. The real value lies between those two limits, in a region that no closed-form formula covers.

Use \(2I_D/V_{ov}\) as a sanity bracket rather than as an answer. If your simulator says \(g_m\) is half or double that, look for a mistake. If it says 27 % less, that’s a 130 nm transistor behaving normally.

\(g_m/I_D\): how designers size devices

Look again at the last column of the \(g_m\) printout, set out here as a table:

\(V_{GS}\) \(I_D\) \(g_m\) \(g_m/I_D\)
0.90 V 63.876 µA 376.932 µS 5.90 /V
1.20 V 217.505 µA 634.008 µS 2.91 /V
1.50 V 436.080 µA 809.693 µS 1.86 /V
1.80 V 696.275 µA 914.650 µS 1.31 /V

Between the first row and the last, the current goes up by a factor of 10.9, but the \(g_m\) you bought with it goes up by a factor of only 2.4. You can think of \(g_m/I_D\) as an exchange rate: it says how much transconductance one ampere of bias current buys you, and it falls steeply as you push the gate harder.

In the other direction, toward lower \(V_{GS}\), it climbs. The Lab 03 deck spice/gm_ro.spice steps the same device down to \(V_{GS} = 0.7\) V:

V_GS = 0.7 V
i(vdc) = -1.06419e-05
@m.xmc.msky130_fd_pr__nfet_01v8[gm] = 1.437123e-04

143.7123 µS from 10.6419 µA works out to 13.50 /V, which is ten times the efficiency of the 1.8 V bias, using a sixty-fifth of the current.

The climb has a ceiling, though. Deep below threshold the drain current is exponential in \(V_{GS}\), and the derivative of an exponential is proportional to the exponential itself, so \(g_m/I_D\) levels off at

\[\left.\frac{g_m}{I_D}\right|_{\max} = \frac{1}{n V_T}\]

Here \(V_T = kT/q = 25.8649\) mV at 27 °C, and \(n\) comes from the subthreshold slope you measured in Lab 03, 85.6 mV/decade, so \(n = 85.6/(2.3026 \times 25.8649) = 1.437\). That puts the ceiling for this device at

\[\frac{1}{1.437 \times 0.0258649} = 26.9\ /\text{V}\]

At \(V_{GS} = 0.7\) V you’re at 13.50, half the ceiling, which is the textbook definition of moderate inversion. That’s the region between the square law and the exponential, where most low-power analog circuits operate and where neither hand formula works.

The single number \(g_m/I_D\) captures the whole design trade. If you bias the device hard, you get speed and small devices, but poor current efficiency. If you bias it softly, you get more \(g_m\) per microamp, at the cost of area and bandwidth. Professional analog designers choose \(g_m/I_D\) first (10 to 15 /V for a low-noise amplifier, 5 for something that has to be fast) and let the geometry follow from it.

\(r_o\) and intrinsic gain

\(g_m\) tells you how much signal current you get for a given input. To turn that current back into a voltage you have to push it through a resistance, and the largest one available is the transistor’s own \(r_o\). It comes from the slope of the saturation region, which looks flat but, as Where saturation starts showed, isn’t quite.

When you multiply the two, the units of current cancel and you’re left with a voltage gain:

\[A_{v,\max} = g_m r_o = \frac{g_m}{g_{ds}}\]

This is the intrinsic gain, the most voltage gain one transistor can produce at a given bias, whatever you build around it. It’s an upper limit because it assumes you loaded the transistor with something infinitely better than itself. spice/gm_ro.spice measures it at four lengths, each at \(V_{GS} = V_{DS}\) = 1.8 V:

\(L\) \(I_D\) \(g_m\) \(r_o = 1/g_{ds}\) \(g_m r_o\) in dB
0.15 µm 2644.190 µA 2468.806 µS 3.588 kΩ 8.86 18.9 dB
0.5 µm 1196.020 µA 1460.824 µS 24.008 kΩ 35.07 30.9 dB
1 µm 696.275 µA 915.312 µS 64.130 kΩ 58.70 35.4 dB
4 µm 197.974 µA 274.342 µS 386.764 kΩ 106.11 40.5 dB

Compare the two ends of the table. The minimum-length device has 9.0× the transconductance of the 4 µm one, yet it gives only one twelfth of the gain, because its output resistance is 108 times lower. Channel-length modulation is the reason: when the drain voltage shortens a short channel slightly, that’s a large fractional change in its length, so \(g_{ds}\) is large.

This is the most important trade in analog IC design, and it runs in the opposite direction to digital design. The digital track, DD101 through DD104, wants every transistor as short as the process allows, because short transistors are fast. An analog designer who needs gain deliberately draws long devices and pays for them in area and speed. Both use the same process and the same transistor, with opposite instincts about its length, and when the two teams share a die, this is what they argue about.

Intrinsic gain in the tens is normal for a modern process. If a hand analysis promises a gain of 500 from one transistor, it used a value of \(\lambda\) from a textbook written when channels were microns long.

You’ll run spice/gm_ro.spice yourself in Lab 04, Part 3.

Where these numbers go next

The inverter. Take an NMOS and a PMOS, tie their gates together and their drains together, and you have two transconductances driving one node, loaded by two output resistances in parallel:

\[A_v = -\left(g_{mn} + g_{mp}\right)\left(r_{on} \parallel r_{op}\right)\]

Both devices push on the output, so their \(g_m\) values add, but their output conductances add as well, so the result is a weighted average of the two devices’ own \(g_m r_o\) rather than a sum. The CMOS inverter measures a gain of −116.0341 at L = 0.5 µm, well above the 35.07 in the table above, because it works at a different bias from the table’s, where each transistor’s own \(g_m r_o\) is about 116. That page has the numbers.

The mirror. A saturated transistor is very nearly a current source, and \(r_o\) tells you how nearly. An L = 1 µm device holding 50 µA changes its current by 1.46 % when its drain moves from 0.9 V to 1.2 V. That device is the heart of the current mirror, and this page’s \(r_o\) column sets the mirror’s error budget.

Check yourself

  1. Why can a transistor have voltage gain when a resistor can’t?
  2. At \(V_{GS}\) = 1.2 V, what does \(2I_D/V_{ov}\) predict for \(g_m\), using \(V_{TH}\) = 0.6016 V? Is the measured value above or below it?
  3. You need a \(g_m\) of 1 mS from as little current as possible. Should you bias near threshold or with a large overdrive? What do you give up?
  4. Going from L = 0.15 µm to L = 4 µm, \(g_m\) falls and \(g_m r_o\) rises. Explain each in one sentence.

Next: Sizing with W and L.