Method for Detecting Noise Parameters of Discrete Semiconductors

Noise Parameter Testing Methods for Discrete Semiconductor Devices

Noise is the parameter nobody thinks about until it kills their signal. A low-noise amplifier with a noisy transistor in the front end picks up everything. A precision voltage reference with a leaky diode drifts for no apparent reason. Noise does not show up in a DC sweep. It does not show up in a gain measurement. It only shows up when you put the device in a real circuit and wonder why the output looks wrong.

Noise parameter testing for discrete semiconductors — BJTs, MOSFETs, JFETs, diodes — is one of the hardest tests to run correctly. The signals are tiny. The environment matters enormously. And one unshielded cable can turn a clean measurement into garbage. This article covers the actual methods, the frequencies that matter, and the setup mistakes that waste your time.

Why Noise Testing Is Harder Than It Looks

Most engineers assume noise testing means hooking up a spectrum analyzer and reading the floor. That gives you the noise floor of your analyzer, not the noise of your device. The device noise in a discrete transistor is often in the nanovolt per root hertz range. Your analyzer noise floor is usually higher. So you need a method that extracts the device noise from the system noise.

The standard approach uses a calibrated noise source and a gain measurement. You measure the device gain at the frequency of interest, then you measure the output noise with the input terminated, then you measure it again with a known noise source applied. From these two measurements, you solve for the device's equivalent input noise. It is not complicated math, but it requires careful execution.

This matters for RF front ends, sensor interfaces, audio preamps, and any circuit where signal-to-noise ratio determines performance. A device with 1dB more noise than expected can degrade your entire system by 3dB or more.

The Noise Parameters That Actually Matter

Not all noise is the same. Different devices have different dominant noise mechanisms, and you need to test the right one for your application.

Thermal Noise and Shot Noise in Diodes and BJTs

Thermal noise comes from the random motion of carriers in any resistive element. Every resistor generates it. Every semiconductor junction generates it. The spectral density is flat across frequency, which is why it is called white noise.

Shot noise comes from the discrete nature of current flow across a junction. Electrons do not flow smoothly. They arrive in packets. This creates current fluctuations that scale with the square root of the DC current.

For diodes, shot noise dominates at low forward bias. Thermal noise dominates at high reverse bias. For BJTs, the base current shot noise and the collector current shot noise both contribute. The base resistance adds thermal noise on top of that.

To measure these, bias the device at the intended operating point. Terminate the input with a known resistance. Measure the output noise spectral density with a spectrum analyzer or a dedicated noise figure analyzer. Convert the output noise to input-referred noise by dividing by the measured gain.

The key frequency range for most discrete devices is 1kHz to 100kHz. Below 1kHz, 1/f noise takes over. Above 100kHz, parasitic capacitance and measurement bandwidth limitations corrupt the data.

Flicker Noise and Its Impact on Low-Frequency Circuits

Flicker noise, also called 1/f noise, rises as frequency drops. It dominates below a corner frequency that depends on the device technology and the operating point. For BJTs, the corner frequency is typically 1kHz to 10kHz. For MOSFETs, it can be as low as 100Hz.

Flicker noise is the enemy of precision DC circuits. A sensor interface that needs 0.1% accuracy at 1Hz cannot tolerate a device with high 1/f noise. The noise integrates over time and creates drift that looks like a slow offset change.

To measure flicker noise, you need a very low-frequency setup. A standard spectrum analyzer does not go low enough. Use a low-noise amplifier to boost the signal, then measure with a spectrum analyzer or a dedicated flicker noise measurement system. Alternatively, use a time-domain method: record the output voltage over a long period, compute the Allan variance, and extract the 1/f noise coefficient from the slope.

Bias the device at the exact current you plan to use. Flicker noise scales with DC current, and a 10% change in bias current can change the noise by 20% or more.

Avalanche Noise in Reverse-Biased Junctions

When a diode or transistor junction breaks down in reverse, it generates avalanche noise. This is broadband, high-amplitude noise that can swamp any signal in the circuit.

Avalanche noise is useful in some applications. It is the basis for hardware random number generators. But in most circuits, it is a failure mode you want to avoid.

To test for avalanche noise, bias the device just below its breakdown voltage. Measure the noise spectral density. If the noise rises sharply as you approach breakdown, the device is entering avalanche. The knee of that curve tells you how much margin you have before avalanche noise becomes a problem.

Do not push the device into full breakdown during this test. You will destroy it. Stay 5% to 10% below the rated breakdown voltage.

Test Setup and Equipment Requirements

The setup determines whether your measurement is real or imaginary. Get this wrong, and no amount of math will save you.

Shielding and Grounding Come First

Noise testing requires a shielded environment. Not a nice-to-have. A requirement.

Use a metal enclosure for the test setup. Ground it at a single point to avoid ground loops. Every cable entering the enclosure should be filtered. Use feedthrough capacitors on all signal lines. Use shielded cables with the shield grounded at the enclosure, not at the instrument.

The device under test should sit on a grounded metal plate. Keep it away from switching power supplies, digital circuits, and anything that radiates. A phone ringing three feet away can inject enough noise to corrupt a low-noise measurement.

This sounds extreme. It is not. When you are measuring nanovolts, everything is extreme.

Choosing the Right Instrument for the Job

A spectrum analyzer works for mid-frequency noise, typically 10kHz to 1GHz. For low-frequency noise below 10kHz, you need a dedicated low-noise amplifier and a narrow-band voltage meter or a digitizer with FFT capability.

For the most accurate results, use a noise figure analyzer or a dedicated noise parameter test system. These instruments have built-in calibration routines that extract device noise from system noise automatically. They cost more than a spectrum analyzer, but they save weeks of debugging.

If you do not have access to a dedicated system, you can build a two-stage method. First, measure the system noise floor with the input terminated. Second, measure the total noise with the device connected and biased. Subtract the system noise in power terms (not in dB). The result is the device noise, assuming the gain is high enough that system noise is negligible. This approximation breaks down when device noise is close to system noise, which is exactly the case for low-noise devices.

Bias Circuit Design for Noise Testing

The bias circuit must be quieter than the device you are testing. If your bias supply has 100nV/rtHz noise and your device has 5nV/rtHz noise, you are measuring the supply, not the device.

Use a battery for DC bias whenever possible. Batteries have virtually no noise. If you must use a supply, add a multi-stage RC filter with a cutoff frequency well below your measurement band. A 10kΩ resistor followed by a 10µF capacitor gives you a 1.6Hz cutoff. Add another stage for better attenuation.

The bias resistor itself generates thermal noise. A 10kΩ resistor at room temperature generates about 13nV/rtHz. If your device noise is 5nV/rtHz, the resistor is louder than the device. Use a lower value resistor if the bias current allows it, or accept that your measurement floor is limited by the bias network.

Frequency-Dependent Noise Behavior

Noise is not a single number. It changes with frequency, and the shape of that change tells you about the device physics.

How Noise Rolls Off at High Frequency

At high frequencies, the device gain drops. The noise at the output drops with it. But the input-referred noise rises because you are dividing by a smaller gain. This creates a characteristic roll-up in the input-referred noise curve.

For BJTs, this roll-up starts around the transition frequency (fT). For MOSFETs, it starts around the frequency where the gate capacitance starts to shunt the signal.

Measure the noise at multiple frequencies from 1kHz to 100MHz. Plot input-referred noise versus frequency on a log-log scale. The flat region is your white noise floor. The rising region at low frequency is your 1/f noise. The rising region at high frequency is your gain roll-off effect.

This plot is more useful than a single number. It tells you exactly where the device is quiet and where it is not.

Temperature Effects on Noise Parameters

Noise increases with temperature. Thermal noise scales with the square root of absolute temperature. Shot noise scales with the square root of current, and current changes with temperature. Flicker noise also increases with temperature, though the exact relationship depends on the device.

Test at the temperature where the device will operate. If it runs hot in your application, test it hot. A noise measurement at 25°C means nothing if the device operates at 100°C.

For temperature-dependent noise characterization, run the test at 25°C, 85°C, and 125°C. Plot noise versus temperature. The slope tells you how much your noise budget will grow as the device heats up. This directly impacts your derating strategy.

Mistakes That Ruin Noise Measurements

The most common mistake is measuring without calibrating. You need to know your system noise floor before you can extract device noise. Run a calibration with the input terminated before every test session.

The second mistake is ignoring bandwidth. Noise power scales with bandwidth. If you change your resolution bandwidth on the spectrum analyzer, your noise reading changes. Always note the bandwidth and normalize to 1Hz when comparing measurements.

The third mistake is DC bias instability. If your bias current drifts by 1%, the shot noise changes by 0.5%. For low-noise devices, that drift can be larger than the noise you are trying to measure. Use a regulated current source with less than 0.1% drift over the measurement period.

The fourth mistake is not accounting for gain uncertainty. The conversion from output noise to input-referred noise requires an accurate gain measurement. If your gain measurement has 1dB error, your noise measurement has 1dB error. Measure gain at the same frequency and bias point as your noise measurement.

Run multiple samples. Noise parameters have process variation. A single device might be quiet. The next one from the same lot might not be. Test at least five devices per lot and use the worst case for your design.

Noise testing demands patience, a clean environment, and the right instruments. Skip any of these, and your data is fiction. Take the time to do it right, and you will catch problems that every other test misses.