[sdiy] Temperature compensation results

[René Schmitz]

Ian,

> In an extreme case, I would imagine that one transistor might have a 
> large defect, such as a dislocation pileup originating from a substrate 
> imperfection, and that the other transistor would not.  Then the Is 
> factors would be *qualitatively* different for the two devices, since 
> one would have to include an extra term to describe the large current 
> from the defect-induced leakage path.

Right, that could happen. Although I think that there would also be a
large mismatch between them caused by Is alone, and on top of that the
temperature dependance of Is is qualitatively altered.

> More generally, I believe a good model would include leakage from a 
> distribution of defects.  Mismatch would then depend on what the actual 
> distribution in each device is.  So I would guess that the boundary you 
> are asking about isn't sharp.

I suspect the same. However the difficulty in that is that you have to
have a look at the chip to get a grip on where the defects are.

> Interestingly, I think this discussion has a connection to some work I 
> did a couple of years ago to select a transistor noise source.  I went 
> through about 50 2N3904 devices looking for the ones with the lowest 1/f 
> noise.  I only found one with really low 1/f noise, and that was the one 
> with the lowest overall noise level.  This suggests to me that most of 
> the devices have significantly more leakage than is ideal.  So I agree 
> that there are non-ideal leakage processes in the pairs we choose as 
> matched.

I guess this is another reason why low noise dual parts have usually
large geometry. The larger the chip, it gets more probable that the
defects are more evenly spread across both halves of the device. (If the
chip gets larger than the average distance between defects that is.)
Also a localised defect doesn't contribute as much as in a small
geometry device.

> Perhaps the devices we pick as being matched have simply have a low 
> level of similar defect distributions.  This suggests a new way to pick 
> matched pairs: compare reverse leakage currents and pick pairs with 
> matched and relatively low leakage.  Of course no one will agree to 
> doing it this way, because it's not how Moog did it.  :-(

I guess a practical way could be to use the extrapolation of the log(Ic)
vs Vb slope as in the GP parameter extraction. (The leakage currents are
very difficult to measure directly.) But do it for several temperatures.


Cheers,
  René

-- 
uzs159 at uni-bonn.de
http://www.uni-bonn.de/~uzs159