In response to our discussion Artto Aurola from Pixpolar commented on the company's MIG operation and benefits over conventional sensors. To get more visibility for his comments, I'm moving them to the front page:
I was asked to comment this conversation about Pixpolar's Modified Internal Gate (MIG) technology. The MIG technology has excellent low light performance due to high fill factor, high sensitivity, low noise and long integration time (allowed by multiple CDS read-out which is enabled due to MIG sensors non-destructive read-out ability). Therefore the low end of the dynamic range can be extended considerably in MIG sensors. A reasonable value for the dynamic range can also be obtained in the high end of the dynamic range scale – typically it is easier to improve the high end than the low end of the dynamic range in any sensor. In case you want to dive deeper into the technological aspects you'll find below a summary about the low light benefits of MIG sensors compared to Complementary metal oxide silicon Image Sensors (CIS) and to the DEpleted Field Effect Transistor [DEPFET, aka Bulk Charge Modulated Device (BCMD) etc] sensors and few words about pixel scaling.
MIG COMPARISION TO CIS IN LOW LIGHT
A critical problem in most digital cameras is the image quality in low light. In order to maximize the low light performance the Signal to Noise Ratio (SNR) needs to be as high as possible, which can be achieved:
1) BY MAXIMIZING CONVERSION AND COLLECTION EFFICIENCY, i.e. by converting as many photons to charges as possible, by collecting these photo-induced charges (signal charges) as efficiently as possible and by transporting these signal charges to the sense node as efficiently as possible. In CIS this is done by incorporating Back Side Illumination (BSI) and micro lenses to the sensor and by the design of the pinned photo diode and the transfer gate. The essentially fully depleted MIG pixel provides throughout the pixel a decent potential gradient for signal charge transport and up to 100 % fill factor in case of BSI. These aspects are missing in CIS but the micro lenses compensate this problem fairly well.
2) BY MINIMIZING NOISE.
2A) Dark noise. The most prominent dark noise component is the interface generated dark current which has been depressed in CIS considerably by using the pinned photo diode structure so that the only possible source of interface dark noise during integration is the transfer gate. In the MIG sensors the interface generated dark current and the signal charges can be completely separated both during integration and read-out. The advantage is, however, not decisive.
2B) Read noise:
2B1) Reset noise. This noise component can be avoided in CIS using Correlated Double Sampling (CDS) which necessitates four transistor CIS pixel architecture. In MIG pixel CDS is always enabled due to the fact that the sense node can be completely depleted.
2B2) Sense node capacitance. In CIS [and in Charge Coupled Devices (CCDs)] an external gate configuration is used for read-out wherein the sense node comprises the floating diffusion, the source follower gate and the wire connecting the two. The signal charge is shared between all the three parts instead of being solely located at the source follower gate meaning that CIS suffers from a large sense node stray capacitance degrading the sensitivity. In MIG sensors the signal charge is collected at a single spot (i.e. at the MIG) which is located inside silicon below the channel of a FET (or below the base of a bipolar transistor) meaning that the sense node capacitance is as small as possible enabling high sensitivity.
2B3) 1/f noise. It is mainly caused by charge trapping and detrapping at the interface of silicon and the gate isolator material of the source follower transistor. It can be reduced in CIS by using CDS and buried channel source follower transistors. However, the problem in the external gate read-out is the fact that the signal charges always modulate the channel threshold potential through the interface. Therefore and due to the large sense node stray capacitance a charge trapped at the interface below the gate of the source follower transistor causes always a larger (false) signal than a real signal charge at the sense node which complicates the read noise reduction in CIS considerably.
In MIG sensors the 1/f noise can likewise be reduced using CDS, i.e. subtracting the results measured when the signal charges are present in the sense node (MIG) and after the sense node has been emptied from signal charges (the sense node can be completely depleted!). A complementary way to reduce the 1/f noise in MIG sensors is to use a buried channel Metal Oxide Silicon (MOS) structure which has a relatively thin, high k gate isolator layer (the thin high k isolator is common in normal MOS transistors) and a relatively deep buried channel. This combination reduces the impact of the interface trapped charge considerably resulting in very low 1/f noise characteristics. In other words the signal charge at the MIG results in a much higher signal than a charge trapped at the interface since a) the gate to silicon distance is much smaller than the gate to channel distance, since b) the external gate is at a fixed potential during the read-out, since c) the sense node stray capacitance is very small and since d) the signal charge is not modulating the channel threshold potential through the interface. Instead of using a MIG transistor having buried channel MOS structure very low 1/f noise characteristics can naturally also be achieved by using JFET or Schottky gate in the MIG transistor.
2C) Multiple CDS read-out / double MIG. The read noise can be further reduced in a double MIG sensor by reading the integrated charge multiple times CDS wise. The multiple CDS read-out of the same signal charges is enabled because the signal charges can be brought back and forth to the sense node (MIG) and because the sense node can be completely depleted, i.e. signal charges do not mix with charges present in the sense node. In CIS and CCDs the external gate sense node can never be depleted and once the signal charges are brought to the sense node they cannot be separated anymore. Therefore multiple CDS read-out is not possible in CIS and CCDs. The multiple CDS read-out is a very efficient way in reducing the read noise since the read noise is reduced by a factor of one over the square root of the read times.
3) BY MAXIMIZING THE INTEGRATION TIME. The integration time is typically limited by subject and/or by camera movements. The maximum possible integration time can be achieved if the signal charge is read as often as possible (e.g. the frame rate is set at a maximal reasonable value by reading signal charge in rolling shutter mode without using a separate integration period) and if all the measured results are stored in memory. In this manner the start and end points of the integration period can be set afterwards according to the actual camera/subject movements which can considerably increase the integration time compared to the case wherein the integration time is set beforehand to a value which enables a decent image success rate. In addition the stored multiple read-out measurement results facilitate the correction of image degradation caused by camera movements (soft ware anti shake) which also enhances the maximum integration time and image quality. Finally the stored multiple read-out results enable different integration times to be used for different image areas increasing considerably the dynamic range and the quality of the image.
If the signal charge is read as often as possible in CIS, if the difference between read-out results at selected start and end points is used as a final measurement result and if the sense node is not reset in between, the final result will comprise a lot of interface generated dark current integrated by the sense node and a lot of 1/f noise. If, on the other hand, the interface generated dark noise is removed CDS wise at all the measurement occasions taking place during the afterwards selected integration period the read noise is added up (this applies also to CCDs). Therefore it may be necessary in CIS to limit the number of the read times to one or to two times at maximum. In this manner, however, the integration time must be set in beforehand and the benefits of multiple read-outs cannot be obtained. Especially the adjusting of the integration time to the subject movements becomes more or less impossible.
The big benefit of the double MIG concept is that for a given pixel one can freely choose an integration time between any two measurement occasions taking place during the afterwards selected start and end points whereby the result is free of interface generated dark noise, whereby read noise is not added up and whereby 1/f noise is considerably reduced due to CDS read-out. (To be precise the read noise is at lowest if the pixel specific start point of integration corresponds to pixel reset.) The first and second points are due to the fact that non destructive read-out and elimination of interface generated dark noise is taking place simultaneously. These two aspects are also true for the single MIG pixel. If in single MIG pixel the 1/f noise is at a low enough level (e.g. buried channel MOS approach) the pixel integration time can also be selected independently from other pixels without degrading the image quality (higher 1/f versus higher integration time) although CDS would not be used.
The afore described means a profound change for especially low light digital photography since instead of setting the integration time beforehand the image can be constructed from a stream of afterwards selected sub images taken at maximum reasonable frame rate. The latter enhances also the ability to correct the image quality software wise.
MIG COMPARISON TO DEPFET/BCMD
In the DEPFET (aka BCMD) sensor signal charges are collected in silicon in an Internal Gate (IG) below the channel of a FET where they modulate the threshold voltage of the FET. In DEPFET the sense node is the IG which can be completely depleted. In addition to that the read-out in DEPFET is non-destructive. The difference between MIGFET and IGFET (DEPFET) is the fact that the current running in the channel of the MIGFET is of the same type than the signal charges in MIG whereas the current running in the IGFET channel and the signal charges in IG are of opposite type.
Process fluctuations (e.g. small mask misalignments etc) and/or poor layout design cause small potential wells for signal charges to be formed in the IG of the DEPFET/IGFET sensor. The threshold potential in the FET channel above the locations of the small potential wells is, however, at maximum (signal charges and current running in the channel are of the opposite type). This results in the problem that signal charges are first collected at locations where the sensor is least sensitive, i.e. read noise is the higher the less there are signal charges to be read which degrades image quality in low light. In other words the IGFET/DEPFET sensor is vulnerable to process fluctuations which may degrade the manufacturing yield to an unsatisfactory level.
Another problem in the IGFET/DEPFET sensor is the fact that normal square IGFETs suffer from serious problems originating from the transistor edges necessitating the use of circular IGFETs. The doughnut shaped IG of the circular transistor increases the capacitance of the IG, i.e. of the sense node, leading to lower sensitivity. This is also likely to impede low light performance. The circular IGFET structure also significantly complicates the signal charge transfer back and forth to the IG meaning that multiple CDS read-out is difficult to perform. Additionally the large size of IG is likely to make the device even more prone for process fluctuations.
The benefit of the MIGFET is that the signal charges are always collected under the location of the channel where the threshold voltage is at highest (signal charges and the current running in the channel are of the same type) which means that MIGFET has tolerance against process fluctuations and that read noise is not increasing when the amount of signal charges decreases. Another benefit is that square MIGFETs do not suffer from edge problems. This fact enables the use of a minimum size MIGFET having a small MIG which enables very high sensitivity. In addition the square MIGFET facilitates considerably the signal charge transfer back and forth to MIG enabling multiple CDS read-out.
The less there are signal charges in the IG the higher the threshold voltage of the IGFET. This is a problem when the read-out is based on current integration since the less there are signal charges in IG the smaller the current through the IGFET and the smaller the amount of the integrated charge. The smaller the amount of the integrated charge the more the SNR is degraded due to high level of shot noise in the amount of integrated charge. In MIG sensors the situation is the opposite way round, i.e. the less there are signal charges in MIG the smaller the threshold voltage of the MIGFET and the bigger the current through the MIGFET, and the less the SNR is affected by shot noise in the amount of integrated charge.
In buried channel circular IGFETs the charges generated at the interface of silicon and gate isolator material are blocked at the origin leading to accumulation of interface generated charges at the interface between source and drain. Although the interface accumulated charge can be removed during reset the interface accumulated charge may increase the noise especially when the amount of signal charge is small and when long integration times are used even though the gate isolator material layer would be fairly thin. The ability to use square MIGFETs enables the use of the buried channel without interface generated charges being accumulated at the silicon and gate isolator material interface (the edges of the square transistor form an exit for the interface generated charges).
Instead of using a circular buried channel IGFET one can use circular JFET or Schottky gate IGFETs which do not suffer from interface charge accumulation. The JFET based IGFET is, however, more complex than other structures (i.e. there are more doping layers present) and since the IGFET is vulnerable to process fluctuations the JFET type IGFET is likely to have the lowest yield. The Schottky gate IGFET is, on the other hand, process wise quite exotic and may be difficult to be incorporated to a CMOS image sensor process.
Probably for afore described reasons the only IGFET structure that has been studied extensively is the circular surface channel IGFET. In surface channel IGFET the external gate cannot be used for row selection since it is necessary to keep the channel open continuously (excluding reset of course), otherwise the IG would be flooded with interface generated dark current. This can be prevented if at least one additional selection transistor is added to the pixel (e.g. the source of the selection transistor is connected to the drain of the IGFET). The downside of the fact that the channel needs to be kept open continuously is that the transfer of signal charges back and forth to IG is even more complicated hindering multiple CDS read-out. In surface channel MIGFET the channel can be closed without the interface generated charges being able to mix with the signal charges in MIG (there is a potential barrier in between). Therefore the external gate of the surface channel MIG can be used for row selection. Additionally multiple CDS is enabled in the surface channel MIGFET since the external gate can be used also for signal charge transfer.
Important aspects concerning the pixel size in today’s BSI image sensors are certainly the manufacturing line width, the SNR and the dynamic range. The manufacturing line width defines the lower limit for the pixel size: it cannot be smaller than the area of the transistors belonging to one pixel. In practice the pixel size is defined by the low light level performance, i.e. by the SNR of the pixel. In both aspects MIG sensors have an advantage since the minimum number of transistors required per pixel is only one, since many features of the MIG transistors enable noise reduction compared to CIS/CCDs and since the integration time can be maximized by setting the integration time afterwards (multiple CDS provided by double MIG, low 1/f noise provided by buried channel MIGFET). A very good analogy for the MIGFET sensor is the DEPFET sensor which has the world record in noise performance being 0.3 electrons even though it suffers from the above listed problems.
The smaller pixel size should not, however, compromise the dynamic range of the pixel. The lower end of the dynamic range scale is already optimized in MIG sensors. The small sense node (i.e. MIG) capacitance naturally somewhat limits the full well capacity of the MIG. Again a good analogy is here the IGFET/DEPFET/BCMD technology – ST microelectronics has recently reported that they achieved better performance in a 1.4 um pixel with IGFET/DEPFET/BCMD architecture than with normal CIS architecture.