Local op-amp rail bypassing: Best practices?

[cvanc]

Hi all-

Let's talk about the small capacitor(s) that are typically used for power rail bypassing of op-amps.  The usual idea is 100nF from B+ to ground and 100nF from B- to ground, with both caps being placed in close proximity to the chip.

I recently read about an alternative to this: Using just one 100nF cap, wired directly from B+ to B-, in other words not connected to ground at all.  The idea is this will avoid injecting rail noise into the ground.  That argument makes some sense to me, especially if board layout is less than ideal, but I have to say I've never actually seen this implementation in anything I've ever worked on.

I've got the usual questions here:  What do we think about this?  Have you ever seen this done in a design, and did it work well?  Are there any particular op-amps that just hate this idea and misbehave?

All comments welcome.  Thanks.

[mvs]

What do we think about this?  Have you ever seen this done in a design, and did it work well?  Are there any particular op-amps that just hate this idea and misbehave?

Dave has used this approach in µCurrent Gold. I guess to limit capacitive load on rail splitter amp.

I am not a fan of this solution, if ground is used as return path for low impedance load.

[Zero999]

It depends on the op-amp and the application. If it's a low current, low frequency application, then supply bypassing is not critical and may even be unnecessary.

[David Hess]

For stability purposes, that only works for a load which terminates into one of the supplies.  If the load terminates into ground, then it can be a problem.  In precision applications, the signal and power grounds might only be connected at one point but this becomes problematical at higher frequencies.

Consider exactly where the return currents from each supply pin to the output have to travel.

https://www.analog.com/media/en/technical-documentation/application-notes/6001142869552014948960492698455131755584673020828AN_345.pdf

[spec]

For stability purposes, that only works for a load which terminates into one of the supplies.  If the load terminates into ground, then it can be a problem.  In precision applications, the signal and power grounds might only be connected at one point but this becomes problematical at higher frequencies.

Consider exactly where the return currents from each supply pin to the output have to travel.

https://www.analog.com/media/en/technical-documentation/application-notes/6001142869552014948960492698455131755584673020828AN_345.pdf

This is the answer, for two main reasons: frequency stability/response, supply line rejection, especially at medium and high frequencies.

Also as DH implies, you, must decouple to a point as close to the opamp as possible but between the input 0V reference point and the 0V line from the PSU. Hash on the supply lines (nearly all opamps have class A/B outputs) feed current through the decoupling capacitors to the 0V line. This hash current generates voltage drops, caused by current flowing through traces/wire, and it is vital that this hash generated voltage does not add to the input signal, and cause all kinds of problems.

Bear in mind that the skin effect increases the impedance of traces/wires and that, even the most lowly opamps ike the LM358, have an open loop gain of 100dB (x100,000 voltage gain) up to 5Hz and still have some voltage gain up to 1MHz.

And if you are designing with one of the video opamps you have to be very strict with decoupling and physical layout in order to get a flat frequency response with no ringing.

[floobydust]

I've seen it used in Fostex audio mixing consoles, 24 track.
It halved the number of capacitors required and keeps the audio ground cleaner because the capacitors are not dumping their currents into it. Two separate grounds (audio signal and power) are best.

As long as there are some capacitors from +/- rail to ground to ensure common-mode noise (common to both Vregs) is looked after, the method works fine.

[spec]

I've seen it used in Fostex audio mixing consoles, 24 track.
It halved the number of capacitors required and keeps the audio ground cleaner because the capacitors are not dumping their currents into it. Two separate grounds (audio signal and power) are best.

Sure, this decoupling approach is used, but the OP asked if it is good practice.

As long as there are some capacitors from +/- rail to ground to ensure common-mode noise (common to both Vregs) is looked after, the method works fine.

Once again sure for that mixer, but it would not be fine for a high-end audio amplifier, not to mention a video amplifier.

[Zero999]

This comes back to the point I made earlier: it depends on the application.

I agree, it's best practice to put the capacitors as need to the op-amp's power rails and where the load is connected (I regret not stressing this earlier) but if it's a low frequency/current/DC application, there's no need to get hung up on it.

[David Hess]

One alternative is to use differential signal paths which reject common mode noise to ground.  Now noise injected into the signal ground is rejected by the common mode rejection of the differential circuits.

Where circuits are single ended, the "single point ground" for each stage can be moved from point to point using transformers, baluns, instrumentation amplifiers, difference amplifiers, and currents instead of voltages.  These can all reject common mode noise between two different points.