[time-nuts] measuring noise of power supplies (was: For those that insist on using switching power supplies)

Attila Kinali attila at kinali.ch
Fri Oct 14 18:53:25 EDT 2016

On Fri, 14 Oct 2016 12:00:08 -0600
"Cube Central" <cubecentral at gmail.com> wrote:

> How would one go about testing power supplies and seeing how noisy they 
> are?  I have the standard suite of tools, an oscilloscope and a little 
> (dangerous) know-how.  I am just not sure what to look for or how to safely 
> hook it up to test.

There are different frequency ranges one looks at when measuring
power supplies:

1) Long term stability over seconds to hours
2) Low frequency noise between 0.1Hz and 10Hz
3) Mid frequency noise between 10Hz and 20MHz
4) High frequency noise beyond 20MHz

The borders of these different ranges of noise are kind of arbitrary
and different people use different values. They are motivated by the
need of the consumer (of power) and by the limitations of different
measurement equipment.

The long term stability is what we usually call "wander" and
is mostly dominated by the thermal stability and aging of the
power supply's voltage reference and its control loop. To measure
it correctly one usually uses a high resolution DMM in the 7.5
or 8.5 digits range. One can also build a homebrew system using
a high resolution ADC (like 32bit delta-sigma converters: AD7177-2,
LTC2508-32, AD1262) and a very stable voltage reference (LTZ1000, LM399).
I'd like to refer to the volt-nuts mailinglist on how to properly
do this, as my knowledge in this area is rather limited.

The low frequency range is what we usually call the 1/f region,
although the long term stability also belongs to it. But unlike
the long term region you don't have to sacrifice a virgin to
get decent measurment data. Jim William's appnote[1] has lots of
details how to measure noise in this region. There are slightly
more modern circuits by Todd Owen/Amit Patel[2] and Gerhard Hoffmann[3].
I recently stumbled over a similar amplifier by Enrico Rubiola
and Franck Lardet-Vieudrin[4]. Both [4] and [5] explain why for
low impedance sources (like power supplies) a BJT input stage
would be a better choice than jFETs and also cover the influence
of temperature on the measurement. [6] gives some additional info
on how to design the differential input stage. 
I wonder how an active offset voltage cancelation scheme for
the differential pair input stage using one of the chopper stabilized
opamps (eg LTC2057) would change the temperature dependence and long
term stability  (aka 1/f^a noise) of the circuit, but I have not seen
any measurements of a system like this yet.

The mid frequency range is mostly influenced by the telecom
noise requirements, which for historical reasons cover the 10Hz
to 20MHz range. It is probably the easiest region to measure
with homebrewn instruments. A decently fast ADC with a low
noise voltage reference (like the LTC6655) are all you need.
Depending on how accurately you want to measure the noise, it makes
sense to further split this range into a lower range up to ~500kHz
and an upper range above 500kHz. The reason is that there are today
several high resolution ADCs available that support sampling rates of
up to 1Msps (and some beyond),eg:
AD7982, 18bit 1Msps
AD7984, 18bit 1.33Msps
AD7960, 18bit 5Msps
LTC2386-18, 18bit 10Msps
LTC2378-20, 20bit 1Msps
LTC2368-24, 24bit 1Msps
These would allow to accurately measure the noise range that is
IMHO most interesting for most applications. Interesting because
a lot of applications are insensitive to noise below 1Hz or even
below 10Hz and noise above several 100kHz becomes easy to filter
out using inductors, ferrit beads and ceramic capacitors. When
choosing an ADC for this range make sure you check the actual
SNR/SFDR performance as it a higher output resolution not necessarily
corresponds to the actual performance delivered. This becomes
especially pronounced when going higher with the sampling rate
to cover the higher noise frequency ranges. Beyond 5-10Msps 16bit
is the best you can get and conversly the SNR is limited to
something around 90dB-95dB.

The high frequency noise, I mentioned above 20MHz, but probably
the limit is more in the 1-10MHz range, is where radiation
becomes interesting. Ie with increasing frequency it becomes
harder and harder to "isolate" electronics against noise
and it becomes necessary to shield it with metal plates.
This range is important for the switched power supplies that
started this thread. These supplies have usually switching
frequencies between a couple of 10kHz to low 100kHz for high
power and AC/DC supplies and goes up to a few MHz for the low
power (where low power is relative and can be several W) low
voltage DC/DC supplies. And as these supplies are switching hard
they produce lots of harmonics. A badly designed supply can easily
produce very noticable spikes at 100MHz. Measuring this type of
noise is probably easiest with a good digital oscilloscope with
a sufficiently high analog bandwidth and sampling rate. Alternatively
specturm analysers are a good choice too. Going homebrew in this
range is kind of difficult, as high resolution ADCs max out around
100-125Msps, with the odd exception of LTC2107 that gives 16bit up
to 210Msps (and still an 80dB SNR). Ie the maximum achievable
bandwidth is 40-90MHz for single ADC configurations. But interfacing
these high-speed ADCs requires an FPGA and thus considerable effort.
An alternative approach is to build a spectrum analyser like setup
with a tunable oscillator and a down-mixer. The problem here would
then be to get a flat frequency response and the required calibration.
Another approach is to use an SDR system with a low lower frequency
limit, but this also requires to kind of calibrate it as these are
not ment for power measurements and thus their frequency response
is not really flat. But they give at least a good indication whether
you have any spikes that stick far out.

One thing I haven't mentioned yet is that you will need to have some
form of load that can be switched. Power supply noise is highly dependent
on the current flow, especially for switched power supplies. But
because the load will also create noise, you probably want to just
have a bank of relay switched power resistors. In case you want to
measure the response on load switches, you should replace the relays
by power transistors (otherwise the bouncing of the relay will
confound the measurement).

			Attila Kinali

[1] "775 Nanovolt Noise Measurement for A Low Noise Voltage Reference",
by Jim williams, Linear AN124, 2009

[2] "Measuring 2nV/sqrt(Hz) Noise and 120dB Supply Rejection
on Linear Regulators", by Todd Owen and Amit Patel, Linear AN159, 2016

[3] "A 220 pV/sqrt(Hz) low noise preamplifier", by Gerhard Hoffman, 2014

[4] "Low Flicker-Noise DC amplifier for 50Ω Sources", by Enrico Rubiola
and Franck Lardet-Vieudrin, 2004

[5] "Some Considerations for the Construction of Low-Noise Amplifiers in
Very Low Frequency Region", by Sikula, Hashiguchi, Ohki, Tacano. 2004

[6] "Some Tips on Making a FETching Discrete Amplifier", by George Alexandrov
and Nathan Carter, Analog Dialog 47-10, 2013

Malek's Law:
        Any simple idea will be worded in the most complicated way.

More information about the time-nuts mailing list