Docs › Reference › railRF

railRF

Power integrity on the board shape you actually drew: DC drop, Z(f) against a target, which capacitors earn their place, and what a re-layout cost.

What railRF is, and the four questions

railRF answers power-integrity questions about the board you drew, not about a rectangle standing in for it. It reads your Gerbers, drill and stackup, finds the copper belonging to one net, and solves it — at DC first, then over frequency.

It is built around four questions, in the order a board designer asks them.

The window opens from Tools › railRF, or by opening a .crail in the project tree. It is one resizable, non-modal window per document, and it can sit behind the workspace while you edit the board.

Three buttons on its title bar get a board into it, and they do different things:

Opening a .crail resolves what the document itself names — its artwork, that artwork's stackup and its part library — so a document opens with its board already on screen. A reference that no longer resolves is reported and does not stop the open: the rails, the ports and the target are the document and are still readable, and the status strip says why the board is not there.

What you provide

Ref. and Return: a layer and a net

The rail card has two rows under the rail, and they answer different questions. Ref. is a layer: the copper the rail's current comes back through, usually the ground plane under it. Return is a net: which of the copper on that layer actually is the return.

A layer alone is not enough, because a real plane layer carries more than one net. A 3.3 V island sitting in an anti-pad on the GND plane is on the reference layer and is not ground; a GND pour on the bottom layer, under a supply pad, is not on the reference layer and is ground. Knowing the return net is what lets railRF leave the island out of the return, and keep the pour out of the rail.

Where the return cannot be settled and the rail has copper of its own on the Ref. layer, the rail is refused rather than solved: taking every piece of that layer as the return would count the rail as its own return, and the drop it reported would look like an ordinary number. The refusal names the Return row.

The parts table

Each row is one part on the selected rail. Selecting a row outlines that part on the board.

• Sort by any column. Click a column header to sort the rows by it, click it again to reverse the order, and click it a third time to go back to the order in the .crail. A ▲ or ▼ after a header shows which column is sorted and in which direction. Number columns (C, ESR, f₀, L) sort by value, not by the text, and a row with no value in the sorted column is always listed last. Sorting only changes what you see: it does not edit the document or make it unsaved.
• Find a part on the board. Double-click a row's location to zoom the board to that part. If the board panel is hidden, it is shown first. A part that reads not placed has no position, so there is nothing to zoom to.
• Resize a column. Drag the thin line at the right edge of a column header. Double-click that line to fit the column to its widest entry, counting every row, including rows scrolled out of view. The widths last until the window is closed and are not saved in the .crail.
• A wide table scrolls sideways. The columns keep their widths and the header scrolls with the rows, so each column stays under its header in a narrow panel.

Top and bottom

On a board with copper on its bottom layer, the table has a side column: top or bottom, the side of the board the part is soldered to. railRF reads the part's pads on that side's copper, so a capacitor on the bottom connects to the bottom-layer copper under it, not to whatever top copper happens to be above it. A board with nothing on its bottom copper shows no side column.

The side starts as the artwork has it: a footprint placed mirrored is on the bottom, which is also how the board exports read it, and the location column says bottom after its coordinates. Pick the other side in the row's drop-down where the artwork does not say — a Gerber board whose footprints were all placed on top, say. The choice applies to the selected rows, and to that part on every rail of the document, since a part is soldered to one side; it is saved in the .crail and never in the part library, because the same part number can be fitted top side on one board and bottom side on the next. It does not change the layout.

The part library

The book button under the parts table, and the parts table's right-click menu, offer two things. Open part library opens the document's library; when it has none (or names one that has since been deleted) this is Create part library…, which creates one seeded with every part number the document names. Use existing library… reuses a library another design already built (below). One row per part number:

A bias curve can be added only to a capacitor row that states its capacitance.

A curve does not have to be typed. Under it, Import curve… reads a supplier's capacitance-versus-DC-bias .csv (or any two columns of bias and capacitance), and Paste curve — or pasting a multi-line table into a curve cell, or Ctrl/Cmd+V over the curve — reads columns copied from a spreadsheet or a datasheet, separated by tabs, commas or spaces. Either one replaces the curve as one undoable edit. The capacitance unit has to be stated, in the header (Capacitance[F], C (µF)) or on each value (4.7uF, 470n): a bare 0.47 could be farads, microfarads or nanofarads. Bias is read in volts unless it says otherwise. A curve whose first point is more than ten times the row's marked value, or under a tenth of it, is refused as a unit error. A supplier export sampled every few tens of millivolts is thinned to the points linear interpolation needs — a 201-row export typically keeps about 14 — and no bias between two kept points reads more than 0.2 % of the curve's largest value away from the supplier's own number. The report above the table says how many points were read and kept, and names anything to check: a file that does not name the part, a first point far from the marked value, a curve that stops below the part's rating.

A curve cell takes its value when you leave it or press Enter, and the point then moves to its place in bias order and stays selected.

A row classed Other has no impedance-at-one-frequency column, on purpose: a datasheet's "220 Ω at 100 MHz" does not say how much of that is resistance and how much is inductance, and any split would be a guess printed as a model. Give the row the part's own Touchstone curve — a two-port file measured series-thru, as supplier tools publish a bead's — or state an R-L on the rail's row.

Import table… on the library's toolbar reads a .csv of parts into it as one undoable edit — save a spreadsheet as .csv first. A part-number column is required; C, the self-resonance and L are read only where the column header states the unit (C (pF), resonance (MHz), L (nH)) or each cell does (100nF). A row whose Value and C columns disagree imports neither, and says so. Supplier tools often find a part only once its trailing packaging code is deleted, so a part number that is a library row's with its end trimmed is matched to that row — when exactly one row fits — and named in the report above the table.

Sharing one library across designs

A team that buys the same part numbers board after board can keep one .crlib — in a shared folder, next to every project — and have every design use it:

• Use existing library… on the book button (or the parts table's right-click menu) picks a .crlib and asks what to do with it. The default, Use It Where It Is (Use It Instead where the design already has a library), names the library in place: nothing is copied, and a part added or corrected there reaches every design that uses it. The alternative is Copy Into Workspace — a new library here, seeded with this document's part numbers and filled from the one you picked, which nothing else will change — or, where the design already has a library, Merge Into it: the picked library's rows are merged in by the rules below, in the library editor, as one undoable edit that you then save. A library already inside this workspace, for a design with none, is used as it is.
• Copy its rows in. Import table… also takes a .crlib. Part numbers this library lacks arrive whole, with their bias curves; for one both libraries have, only the fields this one leaves blank are filled. Where both state a value and they differ, this library's is kept and the report names both. A row's model file is still read from where it sits, beside the other library, and the report names each one. It is one undoable edit, and the copy is saved with this workspace.
• Point this design at it from the import dialog. Its Part library row names a .crlib anywhere, exactly as Use It Where It Is does.

A shared library lives outside the workspace, so Archive Workspace… offers it as a row of its own, ticked by default, together with every model file it names — they keep their places relative to the library, so the recipient's copy resolves them unchanged, and the design is repointed at the archived copy. Untick it to send the design without the library.

Parts the rail runs through

A ferrite bead, a sense resistor or a switch standing in for its on-resistance is not decoupling: the rail runs through it. Right-click its row in the parts table and choose Make series element (Make decoupling (shunt) takes it back; each is one undo step). On a board, its two pads become its terminals, and a part that the board shows with both pads on the rail's copper is offered directly — Add refdes as series element on the same menu, or Add as series under the table.

To edit a series part's model, right-click its row and choose Edit Model Source…, or double-click its model source cell. The dialog holds its DC resistance, and either an R-L or a Touchstone file; each value is applied when you leave its box (or press Enter), the rail re-solves, and every change is one undo step in the railRF window. Every value is the row's own; a blank one takes the part library's row where that row is classed Other, and the watermark says what that is. A file picked in the dialog stays the row's own until you press Save to library: that makes it the part number's Model file in the part library — adding the part number, classed Other, if the library lacks it — so every row with that part number, on any rail and in any design using the library, uses it. The row then takes the file from the library. If the library has other unsaved edits it is left open for you to save, and the row keeps its own file until you do. The table's model source column names which won for each number, so the same bead on four rails is one library row. A part classed Other that is left as a decoupling row is refused at Run, by name.

A rail may run through several series parts. Each one cuts it, and the pieces are sections: the board's Copper view shades each section differently, every observation port reads the impedance of its own section, and each part's DC resistance is its own row of the drop breakdown, carrying the current of every load beyond it. The sections have to form a tree from the source — a chain, or one part feeding two branches. railRF refuses, naming the parts, a loop (two series paths between the same copper, whose current split a lumped model cannot answer), a part with copper around it (it is shorted out), and copper that no chain of series parts connects to the source.

A series part has to lie between the source and a load, with the rail's copper on both of its pads. One that hangs off the rail — a link between a switcher's inductor and the node the source is anchored on, say, which is upstream of the source — is refused by name, saying which end is on the rail. Either anchor the source on the part's far side, if the supply really enters through it, or take the part's row off the rail.

With no artwork, the series rows are a chain in row order, nearest the source first, and each other part and load sits at the far end unless its row names the element it sits behind.

On a series row the table's columns change meaning: the capacitance column shows the part's model (its R-L, or its file), the ESR column shows its DCR, and the self-resonance column reads —, because the part is not a capacitor. The inductance column shows the L of its R-L, or — where its model is a file. A series part has no mounting loop: it is a link between two pieces of the rail, not a branch to the reference plane.

Modelling a 0 Ω link, a jumper, a ferrite bead or an RF choke

Each of these is a series part: make it one first, as described above. Until you do, railRF treats it as a decoupling capacitor, and the parts table reads unresolved in the columns only a capacitor could fill. A series part's model is two things: its DCR and its impedance over frequency. You can state them on the rail's row, in its Edit Model Source… dialog, or once in the part library on a row classed Other, where ESR is the DCR and Model file is the impedance file. A library row cannot state an R-L. Use a Touchstone file there, or state the R-L on each rail's row.

A blank DCR — nothing on the row and no ESR on the part's library row — is read as 0 Ω, and the DC answer notes that it was assumed. The part carries the rail's current, so enter the datasheet DCR wherever it is not negligible. Editing and saving the part library updates an open railRF window straight away.

Q0 — is this rail connected, and what does it cost to get there?

Press Run. You get, for each port, the voltage there and how far below the source it is; and, under it, where the drop went — every element on the path, ranked, in millivolts and in milliohms.

A run solves every rail at once. The bar beside the buttons shows how far through the whole run it is, and Stop abandons it and leaves the last completed answer on screen. The line at the top of the results column names the rail the numbers belong to, which reading they are, and what it cost — +3V3 · Fast · 4.1 ms. Picking another rail in the selector shows that rail's numbers without running again; only its |Z| curve is swept, which takes a moment and never enters Accuracy.

48.698 mV below the source, drawing 350 mA

    22.722 mV   47%    64.919 mOhm   26.462 mm of 0.209 mm BOT copper (126.9 squares)
        21 mV   43%        60 mOhm   the source's own series resistance
     2.023 mV    4%     5.781 mOhm    4.814 mm of 0.438 mm TOP copper (11 squares)
     1.904 mV    4%     5.441 mOhm    3.927 mm of 0.387 mm TOP copper (10.1 squares)
      0.61 mV    1%     1.743 mOhm   the reference return
     0.219 mV    0%     0.627 mOhm   2 parallel vias, 0.3 mm plated over 1.57 mm at 25 um

Expect the copper to be a large term, not a rounding error. On a compact board with thin inner copper — 0.15 to 0.5 mm traces, half-ounce copper, runs far longer than they look — a single supply trace is routinely tens or hundreds of milliohms, which is more than everything else on the board put together. In the run above it is nearly half the budget.

Three other things come out of the same solve.

• Islands. The run says how many regions the rail is, and how many the reference is. "The rail is three regions" on a net you believe is one is the answer to a question you had not asked yet.
• A drop map on the board, so the gradient has a place rather than a number.
• The via current check. Each layer transition is grouped with the barrels that share it, the worst barrel's current compared against a stated limit, and a count given that would clear the flag. The limit's basis travels with it, because the plating thickness is worth a factor of two and the drill file does not carry it: state it in the stackup's via entry or in the document's own settings, and railRF says which it used. Where nobody stated one it falls back to a drill-size table and says that, rather than pricing a barrel nobody measured.

Q1 — does this rail meet its target?

On the Frequency tab of the results column — the |Z| plot is on that tab only — railRF judges each observation port's |Z| against its target, reporting the verdict as a sentence: "Passes by 0.6 dB at its worst, 100 MHz." A violation names its frequency and its margin in decibels.

No curve at all? The card under the plot on the Frequency tab says why — most often a source with no output resistance; see What you provide.

Three things are on the plot besides the curve and the mask.

• The anti-resonances, each attributed: L(C9) against C(C7–C8). A peak you cannot attribute is a peak you cannot fix.
• Your aggressors, as vertical lines at each fundamental and its harmonics.
• The coincidences — where one of those lines lands within a stated fraction of a peak. Neither the peak nor the harmonic is alarming alone; the two together are the finding.

Q2 — which capacitors are actually doing anything?

You have forty decaps because the reference design had forty decaps. Some of them sit two millimetres from a lower-inductance part and contribute nothing.

railRF re-solves the whole sweep once per capacitor and reports what deleting each one would cost:

C10  100 uF bulk    12.1 dB  -> -11.5 dB.  Removing it fails the target outright.
C1   100 nF 0402     1.1 dB  ->  -0.4 dB.  Removing it fails.
...
C13  100 nF 0402     0.4 dB  ->   0.2 dB.  Removing it still passes.

A row reading 0.0 dB is a part whose removal leaves the worst margin exactly where it was — a candidate for deletion. That is not parallelism: removing one of N equal parts in parallel moves the answer by about 8.7/N dB, so no bank of a plausible size is redundant that way. A genuine 0.0 dB row is a part that is out of band where the judging happens — shadowed by a neighbour with less mounting inductance.

Today the way to find this out is to build the board and start removing parts with a soldering iron.

Q3 — did my form factor break it?

A supplier gives you a reference design. You re-shape it to fit an enclosure: same schematic, same BOM, different outline and different placement. Compare… runs both and reports the delta.

The pairing is the part worth knowing about, because it is the part that could quietly compare the wrong things:

• Ports and parts are matched by refdes, and models by part number only where exactly one unmatched row on each side carries that number. Anything else is reported as unmatched, with the count.
• Nothing is matched by proximity. Two ports 0.2 mm apart stay unmatched however close they get; two ports at exactly the same coordinate are the port that did not move.
• Two sweeps on different frequency grids are refused, not interpolated. The resonance search adds points where each board's own peaks are, so two boards come back on two axes as a matter of course — and they differ precisely where the curve is changing fastest. Sweep both sides on one explicit grid instead.
• The delta is a ratio in decibels, not a difference in ohms: a PDN curve crosses three or four decades, so a difference in ohms is mostly a picture of where the curve is big.
• An excursion band ends where the sign changes. A resonance that moved is two findings — worse here, better there — and merging them would put the reported peak at the one frequency where nothing happened.

Fast and Accuracy: two readings of the same copper