Monday, December 29, 2008

Caeleste on 3D Integration Ideas

Bart Dierickx from Caeleste presented his 3D sensor ideas on IEEE Nuclear Scinece Symposium in Dresden, Germany, Sep. 2008. The presentation talks about the new design approaches in 3D world. One might give up pinned photodiode and 4T CDS scheme and needs to look for other ways to suppress kTC noise.

Bart presents an idea how this can be accomplished without resorting to active reset schemes. I'm unable to understand how come that the proposed reset reduces kTC noise, but the eldo transient noise simulation in the inset seems to confirm this:

7 comments:

  1. With respect to Reset2 switch, capacitors Cr and Cfd are in series so kTC noise charge is defined by the smaller cap. From charge conservation, similar (small) noise applies to the bigger cap. However I think Bart needs to double kTCr noise because every reset1 he injects another kTCr to the Cr cap to be transferred (almost completely) to Cfd during charge share. Another (small) noise is coming from Nobu-san reset switch charge...

    ReplyDelete
  2. No need to double in case of subthreshold reset

    ReplyDelete
  3. My opinion is that the switched Cr is essentially equal to a resistor. So, it should deliver a kT*Cfd noise in full after many switching cycles. The size of Cr defines the time constant or a number of cycles to get to the equilibrium state. Once the equilibrium state is achieved, the full kT*Cfd noise builds up. The less number of cycles, the lower the noise, but the bigger the image lag.

    But I'm puzzled to see the improvement in eldo transient noise. It's hard to argue with simulator...

    ReplyDelete
  4. Typically Eldo transient noise doesn't work with transistors in weak inversion or off when using BSIM models as they are inaccurate in this region (this can be easily checked with just a capacitor and a reset transistor). For this specific circuit, I would only trust EKV models, unfortunately they are rarely available from foundries.

    ReplyDelete
  5. There is no transistor in weak inversion in the proposed pixel. So, eldo should give correct results, unless misconfigured. EKV might be acceptable in some situations, but I was told that it does not pass the SEMATECH tests.

    ReplyDelete
  6. Well, I think all transistors goes through weak inversion before being completely off... And that's at this moment that noise calculations are important, not when transistors are on. It says how much noise is frozen onto the capacitor. A similar trick was used by the now defunct Small Camera by slowly turning off the reset transistor.

    ReplyDelete
  7. I understand what you are saying, this could happen if one turns off the RST transistors slowly, so that the time when transistors are in weak inversion mode is comparable with reset transistor RC time constant, there C is Cfd and Cr. In most practical cases, I believe this is not the case.

    Even if we assume that RST is turned off slow enough, BSIM3 weak inversion noise is 40-50% of the real value, depending on noimod setting. It just introduces a scaling factor in all the noise comparison and should not change the improvement factor, or lack of it.

    However, you remark brought me another idea how eldo could be possibly misconfigured to show the improvement. If the maximal noise frequency is set high enough for Cfd*Rds noise bandwidth, but too low for Cr*Rds, eldo would miscalculate Cr noise, while Cfd noise would appear correct. I'm not claiming this the reason, but I'm unable to find another one to explain the simulation results.

    ReplyDelete

All comments are moderated to avoid spam.