High frequency filtering is no different from low frequency, the values are just smaller, and layout is more critical. Which includes common/differential mode conversion, which tends to be the dominant route of escape, and I would wonder if that is the case here too, i.e. the ground loop voltage across the board, and thus between cables connected on opposite sides of the board, dominates over differential noise instead.
The 74279221100 is unusual among ferrite beads, in two respects:
- Well, sort of a zeroth, its impedance is pretty low, which is significant here, but that's not terrifically unusual;
- The impedance is nearly proportional, from below 10MHz, up to 200MHz. This means the Q factor is high, or ESR is low.
- They actually rate it for DC bias current at all (a miracle among ferrite beads, generally), and impedance hardly drops up to the measured maximum:
From:
74279221100 Datasheet WE-MPSB EMI Multilayer Power Suppression Bead | Würth Elektronik
The demo board is rated 5V 15A output, and down to 5.7V input, we expect a comparable maximum input current; they used two in parallel (sneakily, they stacked the two component symbols, hence the dots on the pins, but do show separate designators) ensuring that, over most of the input voltage range, the 5A rating is respected. Even at the lowest input voltage, the rating is not exceeded by too much; and notice the thermal rating is 10.5A, and they simply were not able to test at higher currents due to other limitations -- indeed, we don't see the impedance having dropped much at 5A. So we have a reasonable expectation that they will provide some filtering value over the full operating range.
What is the ferrite bead doing, then?
An impedance of 3Ω at 40MHz for example, is equivalent to 12nH inductance. Or 6nH for two in parallel.
It ain't much -- it's not much different from the impedance of an equivalent length of wire -- but we can conclude it's inductive, because the slope is close to +1.
This is somewhat of a diversion, but it's a fundamental of impedance spectroscopy, and connects importantly with electronic filter networks and real components.
The value of a continuous complex function (here, Z = R + jX is an analytical function of frequency) can only vary in certain ways, given by the Kramers-Kronig relations. This isn't easy to work with analytically (and, I haven't been able to derive results myself, so I won't begin to go into detail with it directly), but the takeaway is that, if impedance is rising proportional to frequency, and the element is a causal, passive one-port, then the real component is zero and we have a constant inductance \$L = \frac{Z}{2 \pi F}\$. (Or perhaps more correctly, not using absolute Z but the local slope; but the local slope will only be pure-inductive over a range, if it is over the full range i.e. -∞ < F < ∞.) If there is loss, then the slope will be less than 1, i.e. \$Z \propto F^{1-\delta}\$ for some small δ > 0, and there will be a corresponding ESR on the order of \$R = \delta X\$. In other words, the loss tangent. (But, again, I'm a bit spotty on this, as the solution to K-K relations for a local power-law approximation, isn't apparent to me. There may be trig or hyp functions involved, which look proportional only at small δ.)
Indeed in the 200-1000MHz range, we see the |Z| flatten out, a bit shallower than a 1:2 slope on the log-log plot, implying skin effect and other core loss effects are dominating up here. This corresponds to the increase in R and decrease in X.
For an exactly 1:2 slope, we have a Warburg or diffusion element, R = X, φ = 45°, etc., which arises often in electronics, as skin effect is a diffusion process. The impedance of electrochemical elements (those mediated by ionic motion: battery, supercapacitor, etc.) also shows this response.
(Actually, supercapacitors might be different for a related reason: the fractal dimension of the activated carbon composing the electrodes. I suppose it should be a double-whammy, even, since ions do indeed have to diffuse through that structure.)
Anyway, diversion aside: they're basically putting a couple nH between capacitors, which themselves hopefully have less (single nH?) inductance to GND, so some filtering value is obtained. It won't be much, I think -- attenuation is limited by the inductor divider given by element ESLs, and a proper filter component would be preferable, but they probably don't need much, either.
A more promising alternative might be one of those solid-metal-link inductors, SLR4040-220 for example. Notice it's inductive up to the same frequency limit (ballpark 100-200MHz), but has about quadruple the value. The main downside I suppose, it's fairly tall at 4mm, so could be hard to fit under a board -- but they have the XEL6060 (6.6mm tall) right there, so that's clearly no problem as built.
Layout
As for layout, again, it's paramount. Looking at the gerbers, they have solid ground on 3 of 4 layers, plenty of stitching vias, and buried (stripline) connections to the regulator. Ground loop voltage near the IC is attenuated strongly as it diffuses into the surrounding plane; the loop voltage at a modest distance is probably pretty small.
The board could probably be made smaller, by placing the regulator towards one side, and collecting the connectors/headers/TPs along the opposite side, with filtering components clustered together in the middle (approximating a point-like structure, minimizing ground loop voltage across it, i.e., between connectors).
For a dev board, size is probably advantageous anyway, so this was not an issue.
High frequency emissions are pretty low in general, from this family of ICs; the switching edges are controlled specially to minimize harmonics, without compromising too much on efficiency. That does still leave full ripple at Fsw, as any buck converter would; having enough capacitance, and CLC(LC..) filtering to bring it down to unnoticeable levels, would be about all that's left (and, those are conducted not radiated frequencies, so they also didn't do anything beyond the 2.7uH choke). I guess it's peculiar they didn't put a ferrite bead on the output side, as it has a similar path as the input (switch node - XAL choke - capacitors), thus subject to a zero in the filter response due to the choke's capacitance. It's also possible they tested with a resistor load only, not a LISN (i.e., to measure output conducted emissions), and missed the need; or that the array of capacitors worked out adequately on this side (low enough total ESL, and as compared to trace inductance), but the input needed a little bit of cleanup still. Hard to say beyond this; you'd have to measure it yourself, and, short of asking the engineers who worked on it (if they remember at all), underlying motivations are pretty speculative.
