Tuesday, June 16, 2015

0.3e- Noise Histogram

Eric Fossum kindly allowed me to post a histogram of his and Jiaju Ma's image sensor that apparently achieves 0.32 e- rms noise on a single CDS read. It was achieved by minimizing the FD node capacitance as discussed in the March 2015 JEDS article:

"A Pump-Gate Jot Device With High Conversion Gain for a Quanta Image Sensor" by Jiaju Ma; Fossum, E.R.

The low readout noise manifests itself in a comb-like structure on the histogram, where the distance between the peaks is equivalent to 1e-. This preliminary result was first presented at IISW 2015 last week:

24 comments:

  1. This slide shows the very first histogram with discernible peaks and valleys that was measured. Better results have since been obtained and hopefully will be published somewhere before long.

    Photoelectron counting is at a watershed moment around the world. SPADs have been doing this in array form for several years now, with arrays getting larger and pixels (jots) getting smaller. Several groups reported sub 0.5e- rms noise at IISW. I believe we are the first to report peaks and valleys in the photoelectron counting histogram (PCH) without the use of avalanche gain. Of course SPADs show this PCH regularly.

    The QIS concept is also at a watershed moment. The two tallest tentpoles are low power readout for millions, if not billions of jots at high field rates, and demonstration of a jot that is sensitive to a single electron with low read noise (below 0.3 e- rms) and scalable to submicron dimensions. At IISW we reported good progress on both fronts. That coupled with the CFA modelling results and the excellent SPAD QIS video imagery shown by the Edinburgh team gets QIS to the starting line.

    I have not been this excited about advancement in image sensors since the 93-95 time frame, which was a watershed moment for CMOS APS (CIS).

    ReplyDelete
    Replies
    1. great achievement! just want to know why several readout counts are at 0 or even negative DN?

      Delete
    2. There should be quite a few counts at zero, depending on the mean, according to the Poisson distribution. Broadening of the peak due to read noise can make the distribution go negative.There may be a few DN of offset as well. You can determine that (and CG) from plotting the peak positions vs DN.

      Delete
    3. Eric, nice work, congratulations.
      with a bit of imagination one can seen this quantization also in IMPACTRON: IEEE TRANSACTIONS ON ELECTRON DEVICES, VOL. 50, NO. 1, JANUARY 2003, but the plot there does not have enough data points to clearly claim such a feature and we have not pursued this any further. Of course, any EMCCD or IMPACTRON should also show this.
      Best regards, Jerry Hynecek

      Delete
    4. Thanks Jerry. I looked at Fig 9 in the paper you mention and perhaps there is some peak and valley action. There is a lot of upside to using avalanche either in the detector like SPAD or for readout like your impactron aka EMC. I only think we show peaks and valleys for the first time WITHOUT the use of avalanche gain. There are several downsides to using avalanche and using a more passive gain such as small FD capacitance probably has scaling advantages. In the meantime, I hope some EMCCD user or manufacturer has published or will soon publish their PCH data just for comparison purposes. I guess their valley to peak ratio should be close to zero, so better than what we have seen so far in the jots we made.

      And in case the readers of this blog don't know, IISS was very pleased to honor Jerry with the Exceptional Lifetime Achievement Award at IISW. Jerry previously won the Walter Kosonocky Award in 2003 for the Impactron (aka EMCCD).

      Delete
    5. I think that noise factor of 1.41 should enlarge the peaks in an EMCCD. It should be interesting to check with an EBCMOS since its noise factor is very close to 1. Maybe Intevac or Photonis friends can give us some hints. Very interesting direction indeed !

      -yang ni

      Delete
    6. First time in a CMOS imager, but in a CCD without EMCCD (using DEPFET) see: Wolfel, S.; Herrmann, S.; Lechner, P.; Lutz, G.; Porro, M.; Richter, R.; Struder, L.; Treis, J., "Sub-Electron noise measurements on RNDR Devices," Nuclear Science Symposium Conference Record, 2006. IEEE , vol.1, no., pp.63,69, Oct. 29 2006-Nov. 1 2006

      Delete
    7. Thanks for bringing this work to our attention. The single CDS read noise they report is 4.6e- rms. Indeed, after several hundred non destructive readouts at -55C, the DEPFET device was able to achieve about the same read noise we see in our device. Fig 12 in the paper is a noise PCH measurement.We will definitely include a reference to this work.

      Delete
  2. The paper shows that "the pump-gate jot with distal FD, tapered RG and smaller SF has the highest
    conversion gain of 380uV/e-", but CG here is 242uV/e-. If two structures are the same? or something makes real CG different from simulation results?

    ReplyDelete
    Replies
    1. There are many design variations we are testing, but the slide is from the very first one that showed the signature peaks and valleys. The simulated CG was close to the measured value. Since the first variation was tested, we have tested others with better results. I just wanted to show the very first one. All measurements thus far are consistent with TCAD simulations.

      Delete
  3. Would like to know what the dark-current of this pixel structure is. Also, how was the inter-metal (routing) capacitance has been reduced; as probably this will limit the lowest junction capacitance? Also interesting would be to know what the source follower loading capacitance was. If the gate geometry was small, does any flicker noise manifest?

    ReplyDelete
    Replies
    1. Hi,
      From our measurement right now, the dark current of this device is lower than 10e-/s, and the detailed results will be included in our later publication. In the paper referred in the post, there is an analysis of FD capacitance components, in which the proportion of inter-metal cap can be found. The inter-metal cap can be reduced mainly by avoiding overlap of metals in the layout. The smaller SF can increase the 1/f noise but also can increase the conversion gain(smaller cap), so it is an interesting trade-off to find the optimized size. We have two variations of SF size, and the 1/f noise result of each will also be included in our later publication. Thanks for your attention.

      Jiaju

      Delete
  4. Hi,
    Why giving that kind of histogram instead of a noise histogram? Is it specific to that kind of sensor?
    By the way, which image or target did you use to obtain such an histo?
    Thanks

    ReplyDelete
    Replies
    1. This "photoelectron counting histogram" or PCH is quite useful when the read noise goes below about 0.5e- rms. For most current CMOS image sensors, it is not so interesting since the peaks and valleys do not appear. There is no target, just low light illumination.

      Delete
  5. Hello,
    As the main 3 peaks in the histogram are merged (no more separate) at ~0.6 of the amplitude (70 on the y-axis), I would say that those 3 main gaussian curves have a width of 1.2 electrons, giving a standard deviation of 0.6 electrons.
    Obviously, I missed something !

    ReplyDelete
    Replies
    1. Maybe you did miss something. There are no merged peaks. The valley is the sum of adjacent Gaussian distributions due to read noise. Try plotting the sum of two adjacent Gaussians (say at x=6 and x=7, each with a std dev. of 0.32) and see what you get. The equation for a normal distribution and std dev sigma is easily found on Wikipedia, for example. A better calculation considers all adjacent peaks,each with an amplitude proportional to the Poisson distribution. The valley- peak modulation is a good way to extract read noise in this regime, I think. We are in the process of publishing on this subject.

      Delete
  6. Wonder what will happen if the histogram was plotted for a larger bin-width ?

    ReplyDelete
    Replies
    1. I think you will mostly lose the peaks and valleys. Just like sampling a time domain signal at lower sampling frequency and with a filter. Is your question more complicated than my interpretation?

      Delete
    2. I am just wondering if the comb-like structure is just an artifact of sampling; for different bin widths different peak-trough structure is obtained. Maybe you need more samples...

      Delete
    3. We should make it a principle to include the bin-size with every histogram by default; similar to temperature for dark-current numbers. Without them, useful information cannot be derived from the data.

      Delete
    4. The slide is clearly labeled with 27.4DN/e-. I think this is a completely sufficient number of bins per electron. I agree that this number needs to be given to assure proper sampling has been performed. I also note that the peaks obey the Poisson distribution. Not as noticeable in the slide shown, but in other data we have taken at lower illumination levels, the fit to peak values is excellent (e.g. mean of 0.5-5.0 electrons). The Poisson peak height signature is quite distinctive. There is no doubt about what we are measuring in our case.

      Delete
  7. Jiaju, do you have measurement histogram with sub-photon level illumination or simply a dark signal histogram. Since you have 10e/s, if you operate at 20frames/s, you should be able to see a dual-peak histogram on your dark signal. Thanks !

    -yang ni

    ReplyDelete
    Replies
    1. Yes, we have histogram results from 0.5e- mean signal up to the 30e- mean range. It is just what you (or me at least) expect. Our dark current is substantially less than 10e-/s at RT making histogramming more tedious. Once we have collected a more comprehensive set of data we will write a regular IEEE JEDS open access paper.

      Delete
  8. Great job! I hope you are nice enough not to terminate all other CISs in 10 years.

    ReplyDelete

All comments are moderated to avoid spam.