The CMOS cameras that major companies such as Andor, Hamamatsu and Photometrics offer currently are all sCMOS, meaning they have all the amplifiers and ADCs built into the CMOS chip. In fact, there are 2 parallel amplifiers (low gain and high gain) and 2 ADCs (for low gain/high gain stitching) for each column. The Correlated-Double-Sampling (CDS) is also done on-chip and is called in-pixel CDS. The advantage of this is that they can achieve high frame rates and low read noise more easily in the camera, yielding lower camera costs and impressive specs on paper. However, the problem of matching thousands of pairs of dual gain amplifiers and ADCs to combine into a seamless single linear dynamic range turns out to be non-uniform, less-than-linear and sometimes not even monotonic. This is in addition to the fact that the 4T pixels inside the sCMOS can be up to 10% non-linear. This requires a lot of data processing inside the camera and despite only 1.8e- read noise, such non-uniformity translates to low SNR in the image. In addition, sCMOS generally has small pixels, and some use micro lenses. So the high QE specification on paper doesn’t necessarily translate to good light collecting power.
All pixels in CCD cameras are read out through one collection well and amplifier, so the image is relatively uniform. However EM (Electron Multiplication) is a statistical gain process (Excess noise factor), which effectively increases the shot noise (photon noise) by a minimum of the square root of 2, which is equivalent to reducing the QE by half. However, the square root of 2 is only the theoretical penalty of EM and in practice the penalty from EM can be worse. In addition, high-speed frame transfer (FT) in EMCCDs is a source of clock induced charge (CIC) noise, which deteriorates the image SNR further. The FT time of the image readout incurs additional costs to the SNR because this period during a frame readout is dead time and has a smearing effect on the image. The higher the target frame rate, the greater the effect of this FT time. At 500 fps, the FT dead/smearing time can reach such a high proportion that the effective QE of a 512x512 sensor drops below 40% in combination with excess noise factor.
DaVinci-2K CMOS imager (DV-2K) is low noise CMOS technology uniquely using inverting high gain pixel amplifiers, while having the output amplifiers and ADCs located outside the chip, in the camera controller. There are 16-readout channels with an individual amplifier and one linear and well characterized 14bit ADC for each channel outside the CMOS chip. This promotes uniformity and linearity in the image, and multi-channel readout helps to achieve high frame rates. Since Pixel Reset can be controlled independently of the Pixel Read, SNR can be further improved by the use of Non-Destructive Read(NDR). In this mode, the image is continuously read out and only reset when necessary. This allows the optimization of the
effective sampling rate and therefore SNR after the data are acquired. However, all these features require a much more sophisticated controller. Two full mode camera-link frame grabbers with custom firmware are used to read out the data at 100fps in full frame CDS mode - 2048x2048, and at 200fps in full frame NDR mode - 2048x2048. Thus, in most situations, DV-2K can acquire twice as many frames as the sCMOS in the same amount of observation time with greater SNR. The DV-2K also has large pixels (15x15 microns), similar to that of an EMCCD, with large light collecting power, but without the EM penalty or frame transfer dead time penalty.
Despite low read noise, both sCMOS and EMCCD suffer from added noise when photons are present, due to the inherent flaw of the on-chip photon amplification (in CCD) or digitization (in sCMOS) scheme. On the other hand, DV-2K uses a classically linear readout channel with a single output amplifier and a single ideal 14-bit ADC per channel outside the CMOS chip to avoid these additional noise penalties. In addition, the Non-Destructive-Read capability of DV-2K allows the increase of SNR many fold over CDS mode for very dim applications.