The new CMOS chips coming onto the market have on-pixel converters, which limits the available bit depth. The ASI1600 for example has only 12 bit conversion and generates 16 bit data by simply padding the bottom 4 bits of the 16 bits with zeros. Theory says that the bottom 4-5 bits from a typical 16 bit camera are dominated by read noise, so they have no value anyway and that 12 bits is enough for all but the highest dynamic range chips - is this true?
To test, took 10 subs from my genuine full 16bit H694 CCD camera and divided the fixed point pixel values in the subs by 16 to yield 12 bit fixed point data. I then stacked the new 12 bit subs and scaled the floating point stack for comparison with a normal stack of the original 16 bit data. No pixel rejection was used and the calibration was not perfect, so the results are a bit ropey, but good enough.
Compare the 12 bit and full 16 bit stacks below (STF stretched) - individual pixel values may be very slightly different, but there is no real change in the depth, detail or noise. Although the H694 has a 16 bit converter, 12 bits is actually enough for it - which is consistent with its ~11 bit dynamic range. Conclusion, there is no loss of depth or detail in using a 12 bit converter for a chip with the 11-12 bit dynamic range of typical CCD or CMOS chips. CCD or CMOS chips with greater than 12 bit dynamic range will need more than 12 bit conversion, but there is no penalty in the 12 bit conversion of the current ASI1600.
Thanks for that test Ray. I certaimly can't 'see' any difference in your photos.
Excuse my poor memory, but i recall in previous readings suggestions that there is a certain number of subs required in any stack to optimise 12bits. Was it 64? Or 74, somewhere around there. I can't recall where i saw it.
I could have posted some measurements, but there was no meaningful difference, so put up the images for those who don't trust numbers.
As far as I know, there is no specific requirement for the number of subs needed at 12 bits, but I think that there was some discussion on CN to that effect. If you stack subs, you will increase the bits covered by the target and reduce the number of bits dominated by noise, but the process gives the same result regardless of how many bits in your original data - provided the original bit depth is enough to cover the dynamic range. Stacking 2x12 bit subs gives a 13 bit result, 4x12 bit subs gives 14 bits, 8x12 bits gives 15 bits etc, so quantisation noise decreases with stacking in the same way that normal noise does.
When you mentioned calibration - was the read noise being subtracted? Wouldn't then be also necessary to account for bits taken up by read noise in a CMOS-acquired 12-bit data, so perhaps scaling should be down to say 7-10 bits (instead of 12-bits), depending on gain setting in a CMOS camera?
When you mentioned calibration - was the read noise being subtracted? Wouldn't then be also necessary to account for bits taken up by read noise in a CMOS-acquired 12-bit data, so perhaps scaling should be down to say 7-10 bits (instead of 12-bits), depending on gain setting in a CMOS camera?
Hi Suavi
Calibration was a standard dark subtraction and flat compensation - read noise remained as is ( actually I would love to find a way to subtract the read noise ). I guess I could have rescaled the raw 12 bit data by padding out with zeros, but didn't see the point - stacking the 12 bit data showed what happened.
Sorry, but I don't understand why you suggest 7-10 bits - the 694 and the 1600 both have over 11 bits dynamic range. All I have done is chop off the bottom 4 bits from some 694 data - to show that it isn't necessary to digitise to 16 bits with an 11 bit DR CCD.
With my QHY22 I get JUST shy of 12 bits of dynamic range so I do agree with your results Ray.
Been doing a bit of thinking about it over the course of the day and from what I can figure out, on single subs, even with 12 bits of dynamic range, a 16 bit output should be slightly superior to a 12 bit. Although it may only have 12 bits of dynamic range and it is largely 4 bits of noise, it is still signal and not just padded out zeros.
Calibration was a standard dark subtraction and flat compensation - read noise remained as is ( actually I would love to find a way to subtract the read noise ). I guess I could have rescaled the raw 12 bit data by padding out with zeros, but didn't see the point - stacking the 12 bit data showed what happened.
Sorry, but I don't understand why you suggest 7-10 bits - the 694 and the 1600 both have over 11 bits dynamic range. All I have done is chop off the bottom 4 bits from some 694 data - to show that it isn't necessary to digitise to 16 bits with an 11 bit DR CCD.
Fair enough - my bad with subtracting read noise I had subtracting master bias in mind, and that would normally add more noise anyway...
The part I do not understand though, is why 1600 has over 11 bits of dynamic range available per sub. Is the read noise really as low as less than 1 bit per sub? Why then these chips saturate so quickly, while a CCD with even smaller pixels and with a greater read noise takes significantly longer to saturate pixels? Sorry for being so thick and thus slow to understand...apparently there are no silly questions, but I somehow feel that way right now...LOL
Fair enough - my bad with subtracting read noise I had subtracting master bias in mind, and that would normally add more noise anyway...
The part I do not understand though, is why 1600 has over 11 bits of dynamic range available per sub. Is the read noise really as low as less than 1 bit per sub? Why then these chips saturate so quickly, while a CCD with even smaller pixels and with a greater read noise takes significantly longer to saturate pixels? Sorry for being so thick and thus slow to understand...
It is all to do with well depth. At a Gain of 76 it has a well depth of about 9000e- which is shallow but also has a RN of about 2.2e-. This gives a 12 bit dynamic range.
An FLI 16803 with a 100,000e- well depth and 8e- RN has a dynamic range of 13.61 bits.
Thanks Colin. But that still does not clarify for me why does one need to stack so many more subs with the CMOS chip, if dynamic range per sub is the same as in a CCD? ICX 834 and ICX 814 both have smaller pixels than the 1600 (and shallow wells), but one can expose for much longer and fewer subs are needed to achieve the same level of detail in astrophotos. That's why my intuition tells me that dynamic range per sub must be lower in the 1600.
I think it has something to do with there actually being digits in the lower parts with a CCD and not just padded with zeros with a 12 bit. That is just my take on it.
you don't actually Have to expose lots of short subs with a 1600 - but you Can if you choose to, without losing SNR - simply because it has such low read noise..
This actually gives the 1600 higher dynamic range after stacking than any other affordable chip that I am aware of.
To put some numbers on it, compare two representative chips:
a CCD with 10e RN and 100,000e well depth
a CMOS with 2e RN and 10,000e well depth
the sky-limited sub length scales with the square of the RN, so the sky-limited subs with the CMOS will only need to be 1/25 as long as those from the CCD - ie, for a given exposure there will be 25x as many of them. Now, the maximum signal for a stack of N CCD subs will be N*100,000e and for 25*N CMOS subs will be N*25*10,000 - ie after stacking, the two systems will provide the same SNR, but the CMOS will give 2.5x the final dynamic range. Put another way, a star could 2.5 times a bright before saturation on the CMOS cf the CCD.
Actually I was thinking of a single sub only, not a stack.
With gain settings to maximise well depth, wouldn't read noise in 1600 go up as well making it about on pair with ICX814/834 utilised in some more expensive cameras? Then, dynamic range per single sub shouldn't be less in the 1600?
For example MLX814 from FLI has 2e read noise, while QSI claims that their camera with ICX834 (smaller pixels than ICX814) has intrinsic noise of 1.6e RMS.
I am wondering whether there is a way of expressing true dynamic range available in a single sub in the 1600 in decibels
With the 1600 and its 12-bit ADC, there's a point at which (IMO) there's nothing gained by going further...at about gain ~75 the dynamic range hits 12 bits. Dropping the gain further would increase the dynamic range but this can't be captured with the 12-bit output.
At unity gain(=139) the read noise is slightly lower...I measured mine at 1.6e compared with ~2e at gain=75.
Actually I was thinking of a single sub only, not a stack.
With gain settings to maximise well depth, wouldn't read noise in 1600 go up as well making it about on pair with ICX814/834 utilised in some more expensive cameras? Then, dynamic range per single sub shouldn't be less in the 1600?
For example MLX814 from FLI has 2e read noise, while QSI claims that their camera with ICX834 (smaller pixels than ICX814) has intrinsic noise of 1.6e RMS.
I am wondering whether there is a way of expressing true dynamic range available in a single sub in the 1600 in decibels
Log2[Well-Depth/RN]
For my QHY22 the well depth is 18,500e- and I have a 4.7e- so Log2[18500/4.7]=11.94 Bits.
Not quite sure how to convert that into dB though.
Actually I was thinking of a single sub only, not a stack.
With gain settings to maximise well depth, wouldn't read noise in 1600 go up as well making it about on pair with ICX814/834 utilised in some more expensive cameras? Then, dynamic range per single sub shouldn't be less in the 1600?
For example MLX814 from FLI has 2e read noise, while QSI claims that their camera with ICX834 (smaller pixels than ICX814) has intrinsic noise of 1.6e RMS.
I would say that, at low gain, the 1600 looks a lot more like a 16mp version of the 694 and has about 12 bits DR - at high gain it has a read noise of about 1e, but the DR is only about 9-10 bits in a single sub. But it doesn't matter much what an individual sub looks like - stacking a few hundred of them can yield a result that is entirely different from a single sub, since each sub contains only a very small fraction of the eventual signal. At high gain you can take a whole lot of very short subs.
there is nothing magic about the 1600 and the other low read noise cameras that you list will be able to take short subs. It is interesting that there suddenly seems to be a realisation by camera makers that low read noise is important.
edit: clearly 12 bits would not be enough for the two cameras that you listed in your post #17, but as was pointed out in the first post of this thread, "CCD or CMOS chips with greater than 12 bit dynamic range will need more than 12 bit conversion, but there is no penalty in the 12 bit conversion of the current ASI1600."
That is truly remarkable chip then - thank you Colin and Ray for your patience in lifting me up from the darkness of my ignorance. Makes me wonder that it can't be long until the so popular KAF 8300 retires completely and that it won't be offered anymore in new astro-cameras...
). I guess I could have rescaled the raw 12 bit data by padding out with zeros, but didn't see the point - stacking the 12 bit data showed what happened.
I had subtracting master bias in mind, and that would normally add more noise anyway...
