Dark Frames & Bias Frames
I think that there is a little confusion in the distinction between 'dark' frames and 'bias' frames, due to the mixing of the two during download. In all cameras, the dark frame is a combination of the actual dark frame (from the pixels), a 'smeared' dark frame from the cumulative dark current of the pixel columns during readout, and the bias offset applied to the A-D converter to prevent 'negative' pixel values. The real dark frame should be very stable and low noise and the bias frame will also be stable and 'noiseless' (ignoring amplifier noise, which is very low). However, the 'smeared' dark frame component is variable with readout time and is much less significant if the readout is fast. This smeared component is what causes the top to bottom brightness gradient in dark frames and images, as the pixels at the bottom of the image spend more time being 'read out' than those at the top and accumulate more smeared dark current.
A 30-second MX5C dark frame (ambient temperature +20C).
The top to bottom gradient is just visible in this image and has
average pixel values running from 1538 at top to 1564 at bottom.
The result of all this is that you should get good dark frames and images so long as you subtract both the actual pixel dark current AND the smeared dark current. Subtracting the bias is much less important, as it only causes a shift in the image brightness and virtually no extra noise. All of the 'artefacts' in the dark frame are actually in the 'smear' component (apart from a few hot pixels) and this is the source of the bright columns and shading. A fast readout via the USB interface greatly reduces all of these effects, but simply subtracting a reasonably well matched dark frame from the image gets rid of most of these artefacts.
Readout Noise
When you capture a bias frame it is a mix of readout noise and thermal noise from the vertical registers (which are thermally noisier than the pixels). To get the readout noise alone, you need to take two bias frames and add a fixed offset to one (say 1000) then subtract the other one from the offset one. The resultant frame is composed of two RMS summed samples of readout noise plus a noiseless offset. Measure the noise in a window on this frame and divide the result by 1.414 to get the single frame readout noise.
If excessive noise is present, then it is usually due to computer induced effects, such as random timing variations. These are very common in Windows programs and difficult to fix completely. The main thing is to minimise the number of background programs which are running.
The USB connection reduced the noise on my test camera by about 50%. This is mainly due to the removal of random timing delays from Windows (the USB hardware does the timing instead).
Bias Frames & HX916 / SXV-H9 Cameras
The 'Bias' consists of a DC offset (essentially noiseless), which is provided by an internal potentiometer, and a noise component, generated by the process of CCD readout. The noise component is partially 'fixed' pattern noise (repeating consistently in all downloaded frames) and partially random electronic noise from the readout process.
Much the larger part of the bias on the Starlight cameras is the fixed DC bias and this can be subtracted by the usual process of offsetting the 'dark point' when image processing. The 'fixed pattern' part is partly clock generated fluctuations of the background, usually seen as faint vertical striations at regular intervals, and brightness gradients due to a slight build-up of thermal charge in the vertical registers during image download. This gradient will often show some vertical streaks where warmer pixels exist in some columns.
The clock patterns are usually constant with temperature etc., but the gradient will be less at low temperatures. The electronic readout noise is completely random and so cannot be subtracted in any useful way.
If you capture a stack of bias frames at a temperature close to your camera operating conditions and then average or median combine them, you will remove most of the readout noise component and the resulting bias frame will correct for both the DC offset and the fixed pattern noise, leaving just the readout noise of the original frame + the image data. If you do not use darks, as is recommended for the HX916 and SXV-H9, the bias subtraction alone will level the background and make it a little smoother. However, the sky 'shot noise' will overcome the bias patterns and DC offset in a short exposure under typical skies, so bias subtraction is mainly useful when stacking a large number of short exposures where the multiple additions of any fixed patterns can make them apparent in the final image stack.