Chipworks published a nice review of recent FSI and BSI pixel implementations by Ray Fontaine. I was particularly impressed by the recent Foveon-Sigma SD1 Merill pixel managing to squeeze 6T 3-color pixel in 5um pitch:
Foveon 15.3 Mp x 3 CIS from Sigma SD1 – Pixel Cross Section SCM, Planar View of Pixel at Poly
I'm not sure about the latest generation, but the previous one was manufactured by Dongbu, Korea. Some time ago Sigma talked about its intentions to move the production to Japan. I do not know what happened with these plans.
What surprised me the most : apparently they use 3 times a 4T-pixel concept with a shared floating diffusion, shared reset, shared source-follower and shared row selector. Very nice design ! Are these pixels published somewhere ? So far, I haven't seen anything of then. BTW, also very well done by Chipworks.
Albert, I see what you mean by 3-way shared 4T but I guess that the transfer gates act like switches (potential is same on both sides) and there is not complete charge transfer and thus kTC noise. If the the transfer gates are just barriers, then they might have lag plus kTC noise. I guess we won't know until someone publishes the noise specs.
None of this addresses one of the other primary problems with silicon-self-filtered color -- the large amount of color processing and poor YSNR10.
Both still seem a large price to pay for eliminating some color aliasing, esp. considering 1.1 um pixels below the diffraction limit (anti-aliasing by physics), BSI QE, and sub 2- read noise.
It is a long way down to those deeper wells and you would need a long neck that would be fully depleted at high potential to reach down there and do complete charge transfer out. Not impossible, just unlikely at these dimensions and operating voltages.
Who is foundry of this stuff?
ReplyDeleteI believe the foundry is somewhere in Korea.
ReplyDeleteI'm not sure about the latest generation, but the previous one was manufactured by Dongbu, Korea. Some time ago Sigma talked about its intentions to move the production to Japan. I do not know what happened with these plans.
ReplyDelete5um@Foveon / 4 = 1.25um@Bayer.
ReplyDeleteSo competitive!
If you like to compare it this way, then 2x2 Bayer equivalent should be 5um/2=2.5um.
ReplyDeleteWhat surprised me the most : apparently they use 3 times a 4T-pixel concept with a shared floating diffusion, shared reset, shared source-follower and shared row selector. Very nice design !
ReplyDeleteAre these pixels published somewhere ? So far, I haven't seen anything of then.
BTW, also very well done by Chipworks.
Albert, I see what you mean by 3-way shared 4T but I guess that the transfer gates act like switches (potential is same on both sides) and there is not complete charge transfer and thus kTC noise. If the the transfer gates are just barriers, then they might have lag plus kTC noise. I guess we won't know until someone publishes the noise specs.
ReplyDeleteNone of this addresses one of the other primary problems with silicon-self-filtered color -- the large amount of color processing and poor YSNR10.
Both still seem a large price to pay for eliminating some color aliasing, esp. considering 1.1 um pixels below the diffraction limit (anti-aliasing by physics), BSI QE, and sub 2- read noise.
Lastly, I concur, nice work by Ray Fontaine!
Does the Foveon sensor have anything to do with these patents:
ReplyDeletehttp://www.google.com/patents/US7745773
http://www.google.com/patents/US7339216
These patents apply to devices with 3.3 and 2.5 micron pixels that Foveon once made for use in cell phone cameras. .
DeleteNope, not at all.
ReplyDeleteEric, why do you think there is no complete charge tranfer please ?
ReplyDelete-yang ni
It is a long way down to those deeper wells and you would need a long neck that would be fully depleted at high potential to reach down there and do complete charge transfer out. Not impossible, just unlikely at these dimensions and operating voltages.
DeleteI think their potential design is so great, because lauching products with vertically aligned photodiode is not so easy work.
ReplyDelete