The Designer's Guide Community Forum

The Designer's Guide Community Forum
https://designers-guide.org/forum/YaBB.pl
Design >> Mixed-Signal Design >> Current sources in the subthreshold region
https://designers-guide.org/forum/YaBB.pl?num=1147255964

Message started by Visjnoe on May 10^th, 2006, 3:12am

Title: Current sources in the subthreshold region
Post by Visjnoe on May 10^th, 2006, 3:12am

Hello,

as an analog/mixed-signal designer, I'm a bit biased (I think) to ensuring that current sources remain in the nice saturation region
over all PVT variations.

I'm wondering if there's something wrong/dangerous about biasing current sources (e.g. current source of dif pair) in subthreshold due to e.g. voltage headroom limitations?

Let's assume all PVT simulations + Monte-Carlo come out nicely, are there still reasons not to do this?

I've heard people stating that transistors in the subthreshold region won't match. When I say that any performance degradation would pop up in Monte Carlo simulations, they claim that these sources of error are not in the foundry transistor models.

Does anyone have more experience/insight on this?

Kind Regards

Peter

Title: Re: Current sources in the subthreshold region
Post by vivkr on May 10^th, 2006, 4:33am

Hi Peter,

I think it is undesirable to place current mirrors in weak inversion. Let me list some reasons:

1. Matching is poor. Current mismatch in current mirrors will get worse with smaller (VGS-VT). In fact,
it should normally be quite poor in weak inversion regime.

2. Noise from the current mirrors will shoot up. By placing transistors in weak inversion region, you are maximizing
Gm/I for these devices. As a result, the noise contribution is significant, especially if the current mirror happens to be
used in an opamp as an active load. Of course, then you would also have to worry about its speed.

3. The mirror will be very slow. It is worth remembering that even if you are simply using the mirror to feed in a DC current,
you do expect the mirror output to stay undisturbed due to changing signals elsewhere or in the circuit which the mirror is feeding.
Otherwise, performance of the whole circuit is affected. Fast circuits need fast-responding bias as well.

4. Finally, your simulations are only as good as your models. While some foundries specializing in analog processes may characterize
the operation in weak inversion well enough, the same in not true for all. Then, the results from simulation are unreliable.

There ought to be a longer list of reasons, and I think other people would have more constructive inputs too.

Regards
Vivek

Title: Re: Current sources in the subthreshold region
Post by ACWWong on May 10^th, 2006, 4:55am

I also like to to have saturated current sources, but due to extremely low currents requirements (in deep submicron process) of the circuits i design, i have been forced to use subthreshold current mirrors.

The kits i have used have predicted the increased MC mismatch with low/no overdrive, but i used l=10u to try to improve mismatch due to area ( but i also used low W/L (<1) to get vgs up to as near threshold as possible ;)).

Anyway we were concerned as to the model accuracy also, but when we got the chips back the measurements supported the monte-carlo anaylsis, so we had the benefit of a good kit (IBM).

So i guess it depends on your kit... if you are getting results from your MC analysis which degrade as you expect then, i guess you have to trust it... if the analysis doesn't do this, then speak with your process vendor and ask pretty please for some better models or do a test chip (unfortunately a test chip is usually less painful/only route in most cases!).

Title: Re: Current sources in the subthreshold region
Post by Visjnoe on May 10^th, 2006, 11:52pm

Hi,

I just would like to note that I'm not sure about the bad matching argument. The rule of thumb that current mirrors require a large Vgs-Vt
for good matching is derived from the (ideal) MOSFET square law, with the transistors in the saturation region. So, you can not make any
statements about the matching of transistors in another operating region (subthreshold), based on this derivation (albeit that you may, in the end, be right that subthreshold operation is bad for matching).

Kind Regards

Peter

Title: Re: Current sources in the subthreshold region
Post by ACWWong on May 11^th, 2006, 12:44am

Hi Peter,

Yes, I am relying on emprical data the fab has measured and fitted to the BSIM model, (which in most cases for modern processes it is hard to make devices obey near square law...). The observation was made during simulation. For use of long channel lengths, when the overdrive is reduced the matching degraded, this continued into the subthreshold region (the modelling doesn't make a clear distinction of strong, moderate, weak saturation, weak inversion as it is based on fitted data, as a designer one makes that distinction).
As I said previously the models i had had measurement data at the sub threshold region, so I could accurately predict the matching performance.
If your models do the same, then you should be able to trust your results.

Cheers

aw

Title: Re: Current sources in the subthreshold region
Post by RobG on May 11^th, 2006, 3:28pm

Just to quickly address the matching issue. If you do the math you get matching that is proportional to 1/(V_GS-V_T). You cannot use this formula in Subthreshold! Let me give you a rough formula assuming that V_T mismatches are dominant (please verify as this is off the top of my head).

σ(ΔI/I) ≈ 2σ(ΔV_T)/V_OD
where
V_OD = V_GS-V_T if V_GS-V_T > 2nU_T
V_OD = 2nU_T if V_GS-V_T < 2nU_T
U_T is the thermal voltage kT/q and n is a process parameter in the neighborhood of 1.3 to 1.8

So... the mismatch doesn't suddenly go through the roof, but if you were to choose an operating point you wouldn't want a mirror biased at ST if matching was important.

Modeling issues may or may not be a big deal - be concerned if you need "absolutes" like gm to be spot on (for AC bandwidth), and especially if you have a circuit that depends on two mosfets operating at different current densities such as the ΔV_GS (constant gm) bias circuit. But things like mirrors and diff pairs work fine in ST.

Leakage can be a very large source of error at hot if you are deep in ST.

The performance versus operating point is a continuim. I wouldn't arbitrarily decide a design is bad because it slips into ST on one corner, but I might wonder why my operating point is changing.

Title: Re: Current sources in the subthreshold region
Post by vivkr on May 11^th, 2006, 11:40pm

Hi Rob,

Yes, it is true that the matching does not go through the roof suddenly on entering WI, as one can argue for a long while about
where exactly SI ends, and WI begins.

However, the matching would be comparable to what one sees in SI with low VGS-VT, and maybe a little worse.

I would not use mirrors in WI as I would not want a bias device with a very large Gm and low speed. Usually, it is possible
to get devices to work in SI even if the currents are of the order of 1uA. You just need longer channel lengths. If one were
pushing the limits and trying to run a circuit with tens of nA, well, that's a different matter altogether.

Regards
Vivek

Title: Re: Current sources in the subthreshold region
Post by RobG on May 12^th, 2006, 2:01pm

Vivkr, I agree that I wouldn't run a mirror in WI given a choice, but making the lengths longer is just going to slow it down even more.... I've been running mirrors and other devices in WI (from nA to mA), when appropriate, for a decade or more and the stigma asociated with WI simply isn't justified. Nothing terrible is going to happen if you have to lower the overdrive to meet some other spec.

Back to the original question, as far as headroom is concerned, keep in mind you still need to keep Vds > ~100mV (at room) even if you are in WI. There isn't much advantage in going deep into WI for headroom.

Title: Re: Current sources in the subthreshold region
Post by vivkr on May 14^th, 2006, 11:46pm

Hi Rob,

I agree with you there. Surely, increasing L too much will kill the speed. So, it is not as it using a very large L to
achieve SI will help on that count, if the current is quite small. However, that is a point on which we can discuss
endlessly.

I am not really prejudiced against WI, although I don't have too much experience of it. Of course,
one often puts opamp diff pairs in WI or moderate inversion to maximize Gm, sometimes without
even realizing.

I would normally beware of using it in active loads as the increased Gm contributes more noise.
However, if you are running with a very small current, and/or have very low VDD, then you more or less
have no choice but to operate in WI.

Regards
Vivek