Camera gain, read noise and ADC bit depth: what e-/ADU and unity gain mean

Three numbers that set the ceiling

Every pixel value a CMOS camera reports passes through an analog-to-digital converter (ADC). The AAVSO photometry guide lists the hard limits: a 12-bit ADC can report at most 4,095 ADU, a 14-bit ADC 16,383 and a 16-bit ADC 65,535. It cannot report negative numbers either, which is why cameras add an offset; a ZWO staff member describes the offset as simply keeping the data from clipping at zero.

The second number is the full well: how many electrons a pixel can hold, as published by the camera maker. The third is the system gain, which AAVSO defines as electrons counted per ADU reported. A gain of 1 e-/ADU counts every electron; a gain of 3 e-/ADU scales three electrons down to one count.

AAVSO describes two ways a pixel saturates: the well fills, or the ADC reaches its top value. Which comes first depends on the gain. Combining the two gives the largest signal a pixel can report (our formulation of AAVSO’s description):

largest reportable signal (e⁻) = lower of: full well, or (2^bits − 1) × system gain (e⁻/ADU)

Unity gain and the gain setting

In a reply on ZWO’s forum, ZWO staff define unity gain as the point where one ADU equals one electron, say it differs between camera models, and describe it as a middle gain suggested for deep-sky imaging.

Keep the two meanings of “gain” apart. The number in the capture software is a control value, not e-/ADU. ZWO’s ASI2600MM Pro manual says the sensor switches to its high-conversion-gain (HCG) mode at a setting of 100 and recommends 0 or 100 for deep sky; on the QHY294 the switch happens at a setting of 1600. Neither number says anything directly about electrons per ADU.

Why a 12-bit camera at unity gain clips near 4,096 electrons

QHYCCD’s QHY294 page gives the example directly: at a system gain of 1 for a 12-bit camera, a pixel saturates at only 4,096 electrons, bright stars saturate easily, and the problem grows with fast optics and long exposures. Its remedy is a lower gain setting, which raises the e-/ADU and so the usable well. The same page explains the opposite pressure: at low gain a sub-16-bit camera works at a few electrons per ADU, and the coarse steps add quantization error. QHY argues that with long exposures or fast optics the photon noise is larger than that error, so it matters less, while for narrowband or short exposures a higher gain lowers both read noise and quantization noise.

Computed for cameras in our data (bits and full well as published):

Camera	ADC	Full well	e-/ADU needed to fit the well	ADC ceiling at 1 e-/ADU
ZWO ASI183MM Pro	12-bit	15 ke	3.66	27.3 % of the well
ZWO ASI533MM Pro	14-bit	50 ke	3.05	32.8 %
QHY294M Pro at its HGC switch point	14-bit	about 14.5 ke	0.89	above the well
ZWO ASI6200MM Pro	16-bit	51.4 ke	0.78	above the well

The 15 ke and 12-bit figures for the ASI183 come from ZWO’s ASI183 manual; the 51.4 ke for the ASI6200 from its product page. The same IMX533 sensor is listed at 50 ke by ZWO, 50–52 ke by ToupTek, 58 ke by QHY and 73 ke by Player One (see the IMX533 camera page), so the e-/ADU needed depends on whose figure you trust. Only the 16-bit camera, and the 14-bit QHY294 at its switch point, can run at unity gain without the ADC cutting the well short.

Read noise as the makers publish it

ZWO’s manuals describe read noise as including pixel diode noise, circuit noise and ADC quantization error, and state that it falls at higher gain. That is why many published figures are ranges spanning the gain settings. Examples from our data, with full lists on the ZWO and QHYCCD camera pages:

• ZWO ASI2600MM Pro: 1.0–3.3 e (manual).
• QHY294M Pro: 1.2–1.6 e in high gain mode, 5.2–6.9 e in low gain mode.
• ZWO ASI6200MM Pro: 0.86 e on the current product page, but 1.5–3.5 e in the v1.4 manual. The product page labels its specs as the newer P25 revision; we record 0.86 e and flag the disagreement.

A single number without a gain or mode is hard to compare. The integration time guide shows where read noise matters for exposure length.

What “16-bit FITS” from a 12- or 14-bit camera means

The FITS standard allows integer pixels of 8, 16, 32 or 64 bits only, so 12- and 14-bit samples travel in 16-bit integers. The file is the same size as from a 16-bit camera: an ASI183MM Pro frame of 5496 × 3672 pixels is 40.37 MB (computed; a packed 12-bit layout would be 30.27 MB, but FITS has none). ZWO’s ASI2600MM Pro manual lists RAW16 output even in the camera’s 12-bit ADC mode. The container is 16-bit; the number of distinct levels per pixel is still set by the ADC. Capture programs may or may not rescale those values, and AAVSO’s quick test for ADC clipping is a pixel at the maximum value for the bit depth, so it helps to know which convention your software writes. The storage calculator uses 2 bytes per pixel either way.

Hardware and software binning

AAVSO’s comparison table states that CMOS cameras bin in software after the ADC, while CCDs typically bin inside the sensor before it. Pre-ADC binning reduces the number of conversions and so the read noise; post-ADC binning, AAVSO says, offers no noise advantage over the unbinned image, though it shrinks the file.

Hainaut’s ESO note shows why: read noises add in quadrature, and reading a 2×2 CCD block once instead of four times gains a factor of 2 in signal-to-noise when read noise dominates. Applying the same rule to four separate reads gives (derived):

read noise of a 2×2 bin after readout = √4 × single-pixel read noise = 2 × read noise

For the ASI2600MM Pro at 1.0 e, a software 2×2 bin carries 2.0 e, where a single CCD-style read would carry 1.0 e.

The “hardware” binning on CMOS cameras is not CCD charge binning either. ZWO’s ASI183 manual says it is done in the digital domain like software binning, uses a 10-bit ADC, and only buys frame rate; ZWO recommends software binning when speed does not matter. The ASI2600MM Pro manual says hardware binning switches its output to 12-bit ADC mode.