High frequency unity gain buffer for driving bondpads (Cl = 25 pF)

[BartasA]

Hi all,

I am looking for a unity gain buffer to buffer analogue signals out of the chip for testing. I need it to operate at least up to 40 MHz and drive Cl = 25 pF with small distortion. I am using 1.8 V devices from TSMC 180nm PDK, but the supply voltage is 1.3 V. Current consumptions does not matter too much, but I am trying to keep it below 1 mA.

I have tried different flipped voltage follower versions, namely class AB version and a super source follower using a quasi-floating gate, but I get very high THD values (40%) [THD function for a transient simulation in Cadence Virtuoso]. Original publications show that it should be well below 1%, although they were simulated/measured at lower frequencies.

Question:
1. What could I be doing wrong to get these THD values? I tried increasing gm, power, but did not see much improvement.
2. Can you propose some good not overly complicated alternatives to achieve my goal?

 

[dick_freebird]

40MHz full power bandwidth and rail-rail swing is no kind of
trivial challenge. I'd start with validating these goals you assert,
and whether they apply to all signals or only a few. If you can
"take most of them off the board" for low swing, low frequency
that will let you focus on the ones that need maximum effort.

Now I question the value of that 25pF load. It may be real. It
may be that you do not want to go with a voltage mode output
driving a capacitive load. Capacitive loading imposes edge rate
degradation which can look like distortion, and output drive
stage will certainly have nonlinearities as you approach the rails.,
requiring feedback to correct (too late).

You might like a 50-ohm system and direct connection to 'scope
channel. If your accuracy can hold up through a V, Z scaling
ladder then you can have 50-ohm ports and a replica voltage
which can be backed out to full on-chip levels (might recommend
a de-embed / cal-map instance, which you can force input and
map output, just to go beyond simulated what-to-expect, how-
to-de-embed). For example a 445/55 ladder would be pretty
close to Z=50 and A=1/10 (you could clean it from there, at
least, with some "test the tester" calibration).