It's not the same as a purely digital example but I once designed a board
that used a direct digital synthesizer (DDS) to generate the reference for
a PLL in a microwave synthesizer. The measured phase noise of the DDS
output was significantly higher than it should have been, which resulted in
the final microwave output being pretty crappy since it was effectively
about 500x the DDS output frequency, so that both the bandwidth and the
magnitude of the noise sidebands were amplified. After scratching my head
for quite a while I came across a statement in the DDS datasheet which I
had not noticed before. There was a particular power pin which they said
should be bypassed TO a particular ground pin. I had bypassed that power
pin but the other side of the bypass cap just went to the GND plane and
that particular ground pin also went separately to the plane. I cut some
traces and tacked on a cap directly between those two pins and voila the
noise was cut by about 10dB in a wide swath of the noise bandwidth.

Another story to illustrate how much of a difference a rather short trace
or piece of wire can make. I've designed several moderate power (about 1kW)
motor drives. In one case, we were measuring current by placing one of
Allegro Microdevices' Hall-effect current sensors in line with the drains
of all three high-side FETs in a three-phase bridge. The sensor's current
sense path was a thick conductor only about 3cm long and specified to have
<100 microOhms resistance. The power plane that one end of the sensor
connected to was well bypassed with bulk capacitance close to the FETs, but
since we needed accurate current measurement on the timescale of a fraction
of a PWM cycle we couldn't bypass the FET drains - we had to allow all of
the supply current to flow through the sensor to the FETs. We ended up with
pretty large voltage spikes on the FET drains due to the inductance of the
sensor combined with the fast switching edges of the FET drain currents. We
anticipated this and had incorporated pads on the board for a snubber. I
computed the optimal snubber values and added them to the board. The
voltage spikes were much less but the snubber components were getting quite
warm! This turned out to be acceptable but it was amazing to see how a
relatively low frequency circuit (20kHz PWM) can be affected by only a few
nanoHenries of inductance and can in fact transfer almost 1% of its power
through the energy stored in that tiny inductance each PWM cycle. Of
course, this is because this circuit had very fast (about 100ns) current
rise and fall times as the FETs switched on and off and the resulting dI/dt
was large enough to produce several volts across just a few nH.



On Mon, Dec 4, 2017 at 1:16 PM, Van Horn, David <
david.vanhorn@backcountryaccess.com> wrote:

> Plenty of good answers, and I'll throw in a "HELL YES" as well.
>
> I saw one instance of a product in production using an Atmel AVR, where
> there is a single pin which is an ADC AREF input only rather than a fully
> implemented I/O pin.
> The application didn't use the ADC at all, and the designer thought he
> didn't need that bypass cap.
> There was a box of boards which had failed production test, which had
> resisted all attempts to repair.
> I added the specified bypass cap and recovered 100% of those boards.
>
> If the data sheet specifies bypass caps, design them in.  If it doesn't,
> design them in anyway, you can always DNP (do not populate) in production=
..
>
> Bypasses are your friend.  Route them well, and don't skimp.   I use X2Y
> caps in critical applications. With any type of bypass cap I route so tha=
t
> power goes THROUGH the capacitor pad on its way to the chip, and the grou=
nd
> side of the cap returns directly to the nearest ground pin on the chip.
> Never a "tee" where the current has the option to go past the cap.
>
> Similarly with crystal loading capacitors, where I route them directly to
> the nearest ground pin on the chip and nothing else touches that trace
> until it joins all the ground pours at the chip ground pin.
> I have seen boards fail FCC testing hard because crystal caps were
> "grounded" into a 100 mil ground track that was about a quarter wavelengt=
h
> long at 400+ MHz.  The "ground" actually worked more like a shunt fed
> antenna. :-P
>
>
>
>
>
> --
> http://www.piclist.com/techref/piclist PIC/SX FAQ & list archive
> View/change your membership options at
> http://mailman.mit.edu/mailman/listinfo/piclist
>
--=20
http://www.piclist.com/techref/piclist PIC/SX FAQ & list archive
View/change your membership options at
http://mailman.mit.edu/mailman/listinfo/piclist
.