Why noise is a hard limit
Up to now in this track, the enemies have been things you could in principle defeat: a bias point that drifts, a phase margin that's too thin, a gain that sags at high frequency. Noise is different. It is not a mistake in your schematic and not a parasitic you forgot to extract — it is the restless thermal motion of charge itself, baked into every resistor and every transistor the moment it sits above absolute zero. You cannot design it away. The best you can ever do is push it down and work above it.
Picture a quiet room. Even with everyone silent, put a sensitive microphone in it and you'll hear a faint hiss — air molecules, the electronics, the world refusing to be perfectly still. Your analog signal lives in exactly such a room. The signal is the voice you're trying to hear; the noise is that ever-present hiss. Make the voice loud and it doesn't matter. But whisper — try to resolve a microvolt-level sensor output — and the hiss is suddenly the whole story. Noise sets the smallest signal you can honestly claim to measure.
There's a second reason noise is special: it's random, so it doesn't add the way signals do. Two independent noise sources don't sum their voltages — they sum their *powers*, and you take the square root to get back to a voltage. That single fact, uncorrelated noise adds in quadrature, governs almost every design decision in this guide. It's why doubling something rarely halves your noise, and why buying the floor down is so expensive.
Thermal noise
Heat is motion. In any resistor at room temperature, the electrons are not sitting politely in line — they're rattling around, colliding, drifting back and forth at random. Each tiny jiggle is a tiny current, and across the resistor's terminals those jiggles show up as a faint, fluctuating voltage. This is thermal noise (also called Johnson–Nyquist noise), and it is the purest example of the floor: it depends only on temperature and resistance, not on the current you push through or the signal you apply.
Its defining feature is that it is white — flat across frequency. There is equal noise power in every hertz of bandwidth, from near-DC up past any frequency your circuit cares about. So we don't quote a single voltage; we quote a spectral density, the noise power per hertz. For a resistor R it is:
# Thermal noise of a resistor (voltage spectral density) # Sv = 4 k T R [ V^2 / Hz ] <-- white, flat vs frequency # # k = 1.38e-23 J/K (Boltzmann) # T = 300 K (room temperature) # Example: R = 1 kohm # Sv = 4 * 1.38e-23 * 300 * 1e3 = 1.66e-17 V^2/Hz # sqrt ~ 4.07 nV / sqrt(Hz)
To get an actual rms voltage, you multiply that density by the bandwidth your circuit lets through and take the square root. Wider bandwidth lets in more total noise — a faster circuit hears more of the hiss, which is the first hint of the trade triangle ahead. Transistors hiss too: a MOSFET in saturation generates channel thermal noise that you can model as a noisy current of density roughly 4kTγ·gm, where gm is the device's transconductance and γ is a fabrication-dependent factor near 2/3 for long devices. Notice gm appears — the very parameter that gives you gain also sets a noise term, which is exactly why noise and performance are entangled.
Flicker (1/f) noise
Thermal noise is flat, democratic, and indifferent to frequency. Flicker noise is its moody cousin: it rises as you go toward DC, with a power spectral density that climbs roughly as 1/f. At high frequencies it's negligible; at low frequencies — fractions of a hertz to kilohertz, depending on the process — it can tower over everything else. If thermal noise is the steady hiss of a quiet room, flicker is the low rumble you only notice when everything else goes still.
Where does it come from? In a MOSFET, it's largely charge carriers getting trapped and released by imperfections at the silicon–oxide interface — each trap captures and frees a carrier on its own slow timescale, and the chorus of all those slow events piles up most heavily at low frequencies. Two practical consequences fall out of that physical picture. First, it's worse in smaller devices: a tiny transistor has fewer carriers and a noisier-looking interface, so a big device averages the flicker down. Second, it depends strongly on process and device flavor.
The frequency where flicker noise crosses below thermal noise is called the 1/f corner. Below it, flicker dominates; above it, thermal rules. This corner is the single most important number for any circuit that must measure slow or DC-like signals — temperature sensors, references, biopotential amplifiers. If your signal lives at 1 Hz and your 1/f corner is at 1 kHz, then flicker, not thermal, is your real enemy, and the whole design strategy shifts toward beating it.
Signal-to-noise ratio
A noise floor is only meaningful relative to the signal sitting on it. That ratio is the signal-to-noise ratio, or SNR — the single number that says how cleanly your signal stands out from the hiss. It's a ratio of *powers*, almost always quoted in decibels, and it is the verdict that every noise decision in the chip ultimately rolls up into.
# Signal-to-noise ratio, in decibels (a POWER ratio) SNR_dB = 10 * log10( P_signal / P_noise ) # = 20 * log10( V_signal_rms / V_noise_rms ) # equivalently, in voltages # # 6 dB ~ factor of 2 in voltage (4x in power) # 60 dB ~ signal 1000x larger than noise (in voltage)
That 6-dB-per-doubling rhythm is exactly why SNR maps so cleanly onto digital resolution. Run the sampling and quantization math for an ideal converter and the best SNR you can get from N bits is a famous straight line:
# Best-case SNR of an ideal N-bit converter (quantization noise only) SNR_dB = 6.02 * N + 1.76 # # N = 8 -> ~50 dB # N = 12 -> ~74 dB # N = 16 -> ~98 dB
Noise vs power vs speed
Now the central bargain. You almost never get lower noise for free — you buy it, and the currency is power and area. These three pull against each other so reliably that it's worth picturing as a triangle: noise, power, and speed, with you allowed to favor any two only by giving ground on the third.
Start with the most common lever: bias current. The thermal noise a transistor refers back to its input falls as you raise its gm, and in the cheapest regime gm rises only with the *square root* of drain current. So to halve your input-referred noise voltage you must lower the noise power by 4x, which means roughly 4x the gm, which means about 16x the current. That brutal exponent is why low-noise front-ends are such current hogs. The floor is for sale, but it's priced in amperes, and the price climbs steeply.
Speed is the third corner. Wider bandwidth is faster, but it lets in more white thermal noise — total integrated noise grows with bandwidth, so a faster amplifier is a noisier one unless you spend more current to push gm back up. And there's an area cost threaded through all of it: bigger devices average down flicker and let you store more charge (lower kT/C), but they're slower to drive and eat silicon. You cannot optimize one corner without watching the other two move. There is no free lunch — only a budget you allocate.
* ngspice: measure the integrated input-referred noise of a stage .ac dec 100 0.1 100Meg ; sweep 0.1 Hz to 100 MHz .noise v(out) vin dec 100 0.1 100Meg .print noise inoise_spectrum onoise_spectrum * read the 1/f corner off inoise_spectrum (rising toward DC), * the flat thermal shelf above it, and the integrated total
Designing for the floor
You can't remove the floor, so good analog design is the craft of spending wisely to lower it and arranging the circuit to live above it. The moves divide cleanly into beating thermal noise and beating flicker, plus topology choices that quietly help both.
- Pin down the band that matters. Integrating noise over 100 MHz when your signal lives in 10 kHz is self-sabotage — band-limit the circuit to your real signal bandwidth so you stop collecting hiss you'll never use.
- Buy down thermal noise with gm: raise bias current (eat the ~square-law cost), or get gm more efficiently with a differential pair and clean current sources so noise isn't wasted on a single-ended stage. Bigger isn't always the point — *efficient* gm is.
- Attack flicker structurally: use large devices (especially PMOS) for the input pair to average 1/f down, and keep the critical gain-setting transistors generous in area.
- For DC-precise paths, cancel flicker outright with a technique. Chopping modulates your signal up above the 1/f corner, amplifies it where only thermal noise lives, then brings it back down — leaving the flicker stranded at the chop frequency to be filtered out. Correlated double sampling (CDS) instead measures the noise once and subtracts it from the signal sample, cancelling the slow-moving flicker and offset.
- Simulate the floor with a .noise analysis before trusting any of it, then design the system with margin so the real, fabricated floor still clears your SNR target across temperature and process.
Step back and the whole guide collapses to one discipline. Noise is a physical floor — thermal everywhere and flat, flicker rising toward DC — and your signal must clear it by enough margin to hit your SNR. You lower the floor by spending current, area, and bandwidth, governed throughout by the square-root tax that makes every gain expensive. The best analog engineers aren't the ones who eliminate noise; they're the ones who know exactly what each decibel costs, and refuse to overpay.