In short: Analog noise is the unavoidable random fluctuation added by devices — chiefly white thermal noise and low-frequency flicker (1/f) noise — and good design means referring all of it to the input, comparing it against the signal to get SNR, and spending power and area only where they buy you the most noise reduction.
Every resistor and transistor you place adds randomness to the signal flowing through it. You cannot delete this noise; you can only understand where it comes from, push it below your signal, and stop making it worse. For a 16-bit ADC, an LNA in front of an antenna, or a biopotential amplifier reading microvolts off the skin, noise is the wall that sets how small a signal you can resolve. This article walks through the two dominant mechanisms, the input-referred bookkeeping trick every analog engineer uses, and how SNR ties it to real specs.
Thermal noise: the price of resistance
Thermal (Johnson–Nyquist) noise comes from the random thermal agitation of charge carriers. It is present in any resistive element even with zero applied bias, and it is "white" — flat power spectral density across frequency. For a resistor R the open-circuit noise voltage spectral density is:
Sv = 4kTR [V²/Hz]
At room temperature (T = 300 K, k = 1.38×10⁻²³ J/K), a 1 kΩ resistor produces about 4 nV/√Hz. Integrate that over a 1 MHz bandwidth and you get roughly 4 µV RMS. The two numbers you must keep straight: spectral density scales with √R, and total RMS noise scales with √(bandwidth). Halving your noise floor costs 4× the current (because transconductance and hence noise improve slowly) or means cutting bandwidth.
Rule of thumb every analog engineer memorizes: a 1 kΩ resistor = ~4 nV/√Hz, a 50 Ω source = ~0.9 nV/√Hz. Scale by √(R/1k) for any other value.
In a MOSFET, channel thermal noise appears as a drain-current source Si = 4kTγgm, where γ is roughly 2/3 for a long-channel device in saturation and climbs higher in short-channel nodes. This is why boosting gm (more current, wider device) lowers input-referred thermal noise.
Flicker (1/f) noise: the low-frequency enemy
Flicker noise rises as frequency falls, with a spectral density proportional to 1/f. In MOSFETs it is caused mainly by carrier trapping and de-trapping at the oxide–silicon interface. Its input-referred voltage density follows:
Sv,1/f = Kf / (Cox · W · L · f)
Two design levers fall straight out of that equation. First, bigger devices are quieter at low frequency — flicker noise drops as 1/(WL), so a large-area input pair is standard in precision op-amps. Second, PMOS is usually quieter than NMOS because holes interact less with interface traps, so low-noise front ends often use PMOS inputs. The frequency where 1/f and thermal contributions are equal is the flicker corner, typically a few kHz to a few hundred kHz depending on the process.
Input-referred noise: one number to compare everything
A real circuit has noise sources scattered through every stage. Referring them to the input means dividing each source by the gain between it and the input, so you can represent the whole chain as a single equivalent noise generator at the terminals — as if the circuit itself were noiseless. This is what lets you compare a sensor's signal directly against the amplifier.
The crucial consequence is the first-stage dominance rule: in a cascade, a later stage's noise is divided by the gain ahead of it, so the input stage usually dominates. That is why you spend your power and area budget on the first amplifier and can afford to be sloppy after a gain of 100. (This is the exact dual of the Friis formula for noise figure in RF chains.)
SNR: turning noise into a spec
Signal-to-noise ratio is the ratio of signal power to total noise power, almost always quoted in dB: SNR = 10·log₁₀(Psignal/Pnoise) = 20·log₁₀(Vsignal,RMS/Vnoise,RMS). For data converters it sets the usable resolution through the classic ideal relation SNR ≈ 6.02·N + 1.76 dB for an N-bit converter driven by a full-scale sine. A 12-bit ADC tops out near 74 dB; if your analog front end only delivers a 68 dB SNR, you have effectively thrown away a full bit — your ENOB (effective number of bits) is below 12.
| Mechanism | Spectral shape | Physical cause | Main design lever |
|---|---|---|---|
| Thermal | White (flat) | Carrier agitation in R, channel | ↑ gm / ↓ R / ↓ bandwidth |
| Flicker (1/f) | Rises at low f | Oxide interface trapping | ↑ device area (WL), PMOS input, chopping |
| Shot | White | Discrete carriers over a junction | ↓ DC current (where allowed) |
For a precision DC measurement, chopper stabilization or correlated double sampling modulates the signal above the flicker corner, amplifies it where only thermal noise remains, then demodulates — effectively deleting 1/f noise from the band of interest.
How this shows up in the EDA flow
You rarely compute these integrals by hand on a real block. In Cadence Spectre or Synopsys HSPICE you run a .noise analysis that sweeps frequency, reports the output and input-referred noise densities, and lists per-device contributions so you can see which transistor is hurting you. A transient noise simulation then shows the time-domain jitter or RMS wander. Siemens AFS / Eldo offers the same, and layout-aware extraction matters because a noisy bias line coupling in is often worse than any intrinsic device noise. Noise analysis also interacts tightly with stability — a marginal loop can peak the noise at the resonant frequency, which is why you always check it alongside your phase margin and stability and your op-amp offset and CMRR budgets together.
Mastering this lets you answer the interview classic — "your front end has 68 dB SNR, where would you look first?" — with a real plan instead of a shrug. For hands-on labs on exactly these simulations, teams at Avecas analog design services run the same flows on production silicon.
Want to run these noise simulations yourself on a real op-amp and ADC front end? The CourseTron Analog IC Design track walks you from Johnson noise hand-calcs to chopper-stabilized front ends in Cadence Spectre — with graded lab reports.
Frequently asked questions
Why is thermal noise called "white"?
Because its power spectral density is flat — equal power in every equal-width frequency band, like white light containing all colors equally. It only has a 4kTR value per Hz, so the total you see depends entirely on the bandwidth you integrate over.
Can I reduce noise just by adding gain?
No. Gain multiplies the signal and the input-referred noise by the same factor, so the SNR at the input is unchanged. Gain only helps by making the input stage dominate so that later stages add negligible noise — it never improves the fundamental noise floor of the first stage.
What is the flicker corner and why do I care?
It is the frequency where flicker (1/f) noise density equals thermal noise density. Below it, 1/f dominates. For DC and audio-band precision circuits you want the corner as low as possible, which drives large input devices and techniques like chopping.
How does SNR relate to ENOB in a data converter?
ENOB is derived from the measured SNR (strictly SINAD) using ENOB = (SINAD − 1.76)/6.02. If analog noise drags SNR below the ideal 6.02N + 1.76 dB, your effective resolution falls below the nameplate bit count.