Read Noise

Read noise is the noise floor a sensor exhibits at zero integration time — the irreducible noise the readout chain itself contributes before a single photon's worth of signal has accumulated, measured in electrons rms, and it is the hard lower bound below which no signal, no matter how carefully everything else in the system is optimized, can ever be reliably distinguished from the sensor's own electronics.

Read noise is not a single error source; it is the quadrature sum of nearly every noise mechanism this series has already covered individually, converging into one specification number. Residual kT/C noise survives in the signal chain even after correlated double sampling has done its job, specifically the uncorrelated component from a fresh reset event that CDS cannot cancel. The source-follower transistor that buffers each pixel's signal out of the array contributes its own thermal and 1/f noise, the latter only partially suppressed if the readout timing keeps the relevant frequencies below the flicker-noise corner. The analog-to-digital converter contributes its own quantization noise on top of everything analog that came before it. None of these sources know about each other, so they combine the way independent random variables always combine — added in power, the square root of the sum of squares, not added directly.

This quadrature combination is exactly why read-noise optimization chases the single dominant term rather than treating every contributor as equally important. Because independent noise sources add in power rather than amplitude, a term that is already small relative to the others barely moves the total even if halved completely, while the single largest term dominates the sum almost by itself. A design with kT/C noise contributing the lion's share of read noise variance gains far more from a larger sampling capacitance than from further source-follower noise optimization, even though both are legitimate, real contributors — this is the practical reason read-noise datasheets and optimization efforts tend to name one specific limiting mechanism rather than listing every contributor as equally worth chasing.

Read noise and dark current shot noise are the two fundamental floors that together set where a sensor's useful operating range actually sits, and they dominate at opposite ends of the exposure-time axis. Read noise is essentially constant regardless of how long a pixel integrates, since it is contributed entirely by the readout electronics rather than by anything accumulating during exposure — a short exposure's total noise is read-noise-dominated almost by definition, since there has been no time for dark current to accumulate meaningfully. Dark current shot noise instead grows with the square root of integration time, following directly from the Poisson statistics already established for dark current itself, which means a long enough exposure eventually produces more total noise from accumulated dark current fluctuation than the readout chain ever contributed.

The crossover exposure time between these two regimes is a genuinely useful design and operating parameter, not just an academic curiosity. Below the crossover, reducing read noise — a better source follower, a larger sampling capacitance, a more careful CDS implementation — is the lever that actually moves total system noise, since dark shot noise is still negligible at that exposure duration. Above the crossover, the same read-noise improvements buy almost nothing, because dark current shot noise has already grown to dominate the total; the only levers that matter there are cooling the sensor to suppress the underlying dark current, or simply shortening the exposure back below the crossover if the application can tolerate it.

This is the specific reason cooled, scientific-grade image sensors pair aggressive cooling with carefully engineered low-noise readout chains rather than optimizing either one alone. A sensor cooled to suppress dark current but still saddled with a noisy readout chain only pushes the crossover exposure time further out — useful for long exposures, but it does nothing for a short-exposure application that was already read-noise-limited before cooling was ever applied. Conversely, an exceptionally quiet readout chain on an uncooled sensor still runs into the dark-current shot-noise wall at whatever exposure time the uncooled dark current level dictates. Getting both floors low simultaneously, not just one, is what actually extends a sensor's useful operating range across the widest span of exposure times.

Read noise is where nearly every noise mechanism this series has covered converges into a single number a datasheet can actually publish — kT/C noise, flicker noise, CDS's own limitations, and the readout chain's quantization all feed into it, combined the way independent random processes always combine rather than simply summed. It sits alongside dark current shot noise as the second of two floors that together, not separately, determine how much dynamic range a sensor actually delivers across the full range of exposure times a real application needs to cover.