Skip to content

P8X 8-Slot Backplane — Design Notes

P8X 8-Slot Backplane — Design Notes, KiCad 3D render

The board as routed in KiCad (3D render; top view).

Theory of operation: p8x-backplane-theory.md — deep walkthrough of the bus map, signal integrity, and power.

Built version: the go-forward board is the KiCad build in kicad/ — 8 slots, 28 mm pitch, 262 × 128 mm, routed at 0.13 mm clearance, nylon-screw mounting. The board figures in §2–§4 were brought up to the 8-slot KiCad build on 2026-10-21 (they were the Eagle rev-C notes: 25.4 mm pitch, 10 slots); the bus pinout and theory are unchanged.

1. Bus Pinout

The authoritative human-readable bus pinout is in p8x-bus-definition.md, which is generated from the same busnet() function as the Eagle CAD files. The machine-authoritative source is generators/gen_eagle.py.

Summary for quick reference (rev C2):

Pin Row A Row B Row C
1–2 +5V +5V +5V
3–10 D0–D7 GND/SPARE (alt) A0–A7
11 -RES GND/SPARE (alt) A8
12–15 DOE0–3 GND/SPARE (alt) A9–A12
16–19 DLD0–3 GND/SPARE (alt) A13–A15, ALUS0
20–21 PSEL0–1 GND/SPARE (alt) ALUS1–2
22–23 PINC, PDEC GND/SPARE (alt) ALUS3, ALUM
24–25 CLK, CLKB GND/SPARE (alt) CIN, SH0
26–27 LDF, FC GND/SPARE, CLRC SH1, PSEL2
28–30 FZ, FN, FV BSEL, IRQ, SPARE11 LDZN, SHCIN, SETC
31–32 GND GND GND

A27–A30 = FC/FZ/FN/FV (flag lines; were SPARE0–3 in rev C1). rev C3: C27–C30 = PSEL2/LDZN/SHCIN/SETC and B27 = CLRC (were SPARE4–8). rev C: B28 = BSEL (ALU B-input mux select, was SPARE9), B29 = IRQ (interrupt request, was SPARE10); SPARE11 remains on B30. rev D: B3–B26 alternate — odd pins stay GND (guard between signal pairs), even pins (B4,B6,…,B26) are spare bus lines SPARE12–SPARE23, routed slot-to-slot for future expansion without a backplane re-spin. See p8x-bus-definition.md for the authoritative per-pin map.

Signal positions are unchanged from rev B wherever they existed, so only the power pins moved. Rev B cards are incompatible (rev B grounded B2; rev C puts +5V there). The memory card schematic has been regenerated as rev C.

Current budget: DIN 41612 pins are rated ~2 A each → 6 parallel +5V pins and ~18 ground pins (the 6 pins on rows 31–32 plus the 12 odd-pin B-row guards; rev D reassigned the 12 even B-row pins from GND to spare lines) carry a 3–4 A system with comfortable margin and low connector drop.

2. PCB Construction — this is where the crosstalk battle is won

Use a 4-layer board. At today's fab prices the upcharge is small and it buys more than any other single decision:

Layer Content
L1 (top) Signal: row A nets
L2 Solid GND plane — no splits, no routing
L3 Solid +5V plane
L4 (bottom) Signal: row C nets

This answers the "thicker power traces" requirement in the strongest possible way: the power distribution is two solid copper sheets (use 1 oz minimum, 2 oz if offered cheaply). It also gives every signal a tight return path, which is the primary crosstalk mechanism on a backplane — crosstalk is mostly a shared-return-inductance problem, and a plane under every trace mostly dissolves it.

Layout rules: - Slot pitch 28 mm, 8 slots → board 262 × 128 mm (≈ 20 cm of bus between the end slots) - Bus traces run straight connector-to-connector on the outer layers. The generated board uses 0.2 mm tracks, 0.13 mm minimum clearance (Freerouting), plus a 0.35 mm CLOCK net class for CLK/CLKB (below). Most parallel runs sit at the 2.54 mm pin pitch, where coupling is negligible; some sampled-signal pairs close to 0.33–0.5 mm centre-to-centre for 10–17 cm (measured 2026-10-21, §2.1) - Do not route signals between the planes or split the planes. Spares route like signals (the generated board keeps every track on L1/L4) - CLK and CLKB get the royal treatment: route them with a ground trace on each side (guard traces) or with one empty channel between them and neighbors; keep them on one layer end to end. As built (2026-10-21): a CLOCK net class at 0.35 mm clearance (gen_backplane.py) keeps every neighbour ≥ 1.23 mm centre-to-centre (it was CIN at 0.44 mm, 75 mm alongside CLKB), with both clocks still the ≈ 230 mm daisy chain. Wider is not possible: the DIN pin field leaves 0.395 mm per side for a track threading two pins, and above that the clocks detour round the connector ends. A routed ground guard is beyond the autorouter; the spacing does its job - Decoupling: one 100 nF within 10 mm of each slot's power pins (C1–C8 in the schematic), 2 × 470 µF bulk near the power entry (C11/C12) - Power entry (J11, 4-pos screw terminal: 2 × +5V, 2 × GND) at one end; via-stitch generously into the planes (≥ 8 vias per terminal) - Tie every ground pin straight into L2 at every slot — 18 per slot (the 12 odd row-B guards plus rows 31–32), 144 along the bus, a low-inductance grid - Power LED (R1 + LED1) so a dead PSU is diagnosed from across the room

If you must do 2-layer: top = signals + 10 mm power rails along each edge, bottom = ground pour flooded around a minimal number of crossing traces. It will work at 2 MHz. The 4-layer board will work at 8 MHz. Choose accordingly.

2.1 Dielectric — order a thin outer prepreg

The generator does not set a stackup, so the fab's default 4-layer build applies (commonly ≈ 0.2 mm from L1 to L2). The trace-to-plane height H is the biggest remaining crosstalk lever. A common rule of thumb (Johnson) puts the near-end coupling coefficient near 1 / (1 + (D/H)²), with D the centre-to-centre spacing. It is an upper bound, reached when the coupled run is long compared with the edge; the 10–17 cm runs here, slowed by card loading (§3), come close to it.

Spacing on the routed board H = 0.2 mm (typical default) H = 0.1 mm
2.54 mm (pin pitch, most runs) ≈ 0.6 % ≈ 0.2 %
0.45 mm (tightest pairs) ≈ 16 % ≈ 5 %

So the thin dielectric cuts the worst pairs about 3×, at no layer cost. When ordering, choose a 4-layer stackup with ≈ 0.1 mm outer prepreg (most fabs list one, often with 3313 glass). The 0.2 mm tracks then drop from ≈ 66 Ω to ≈ 41 Ω (microstrip, εr 4.3), which is harmless: the loaded bus sits lower than either (§3). The two levers compound on the clocks: at their 1.23 mm spacing (§2) the estimate is ≈ 2.6 % at 0.2 mm prepreg and ≈ 0.7 % at 0.1 mm. The tight pairs left are sampled signals (crosstalk only has to settle before the clock edge) plus two active-low lines that idle high: -RES (104 mm beside D7 at 0.49 mm) and IRQ (beside FZ/FV/LDZN at ≈ 0.5 mm). Both are safe for the same reason: HCT inputs switch at ≈ 1.4 V, so a line sitting high needs a 72 % dip (5 V → 1.4 V) to see a false edge, against ≈ 14 % (≈ 0.7 V) worst-case coupling. The vulnerable state is a line sitting LOW (a 28 % rise reaches threshold), which is the clock case the CLOCK class protects.

2.2 The connector pin field

At HCT edge rates the connector, not the PCB, is usually the dominant crosstalk path on a backplane. Adding layers cannot change it; only the pinout can.

  • Row B gives every row-A and row-C pin a nearby ground return, but neighbours within row A or row C are still signal beside signal. That coupling lives in the pin field.
  • Keep the plane webs intact. Every DIN 41612 pin is a through-hole barrel that cuts an antipad in every plane. At 2.54 mm pitch, keep antipads ≤ ≈ 1.6 mm for the 1.0 mm drill, so that copper stays connected between pins. Return current concentrates in the connector field.
  • Reserve the spares next to the clocks for slow signals. SPARE22 (B24) sits directly between CLK (A24) and CIN (C24); SPARE23 (B26) is diagonal to CLKB (A25). Assign only sampled, non-edge signals to those two if they are ever used.

2.3 Reference — the stackup with no layer budget

For the record, the low-crosstalk ideal if layer count were free (8 layers, every signal layer against a solid ground):

Layer Role Contents
L1 GND Solid pour and shield; only connector pins pass through
L2 Signal (stripline) CLK, CLKB, -RES, IRQ, strobes: the edge-sensitive lines, widely spaced
L3 GND Solid
L4 Signal (stripline) A0–A15
L5 GND Solid
L6 Signal D0–D7, ALU/flag/control fields
L7 +5V ≤ 0.1 mm to L8: the pair acts as built-in decoupling
L8 GND Solid

Stripline (a trace between two planes) is shielded on both sides. Its far-end crosstalk is near zero, whereas on an outer layer it keeps growing with parallel length. No two signal layers are adjacent.

For this machine it buys little. At 74HCT edge rates the 4-layer board with a thin outer prepreg (§2.1) captures most of the benefit, and the connector field (§2.2) is unaffected either way. For a prototype run an 8-layer board of this size costs several times the 4-layer price (JLCPCB estimate, 2026-10-21).

3. Termination — analysis and recommendation

The physics: the bus is ≈ 20 cm between the end slots (8 slots at 28 mm). Unloaded propagation is ≈ 1.2 ns end to end (≈ 6 ps/mm on an outer layer). Each filled slot hangs ≈ 10–15 pF on every line (connector, stub, transceiver pin); spread along the bus that gives a loaded impedance Z₀ / √(1 + C_load / C_line) of roughly 20–35 Ω (lower as more slots fill) and stretches the delay by 1.4–2.3×. Round trip ≈ 4–6 ns. 74HCT edges are 3–5 ns — comparable to the round trip, which puts the bus at the edge of the transmission-line regime: you'll see some ringing on a scope, but at 2–4 MHz there are hundreds of nanoseconds for it to settle before anything is clocked. Termination is not required for functional correctness at the design speed. Which signals care:

  • Edge-sensitive — anything used as an edge: CLK, CLKB, -RES, latch strobes. One glitch across threshold is a false clock.
  • Sampled — data, address, the encoded DOE/DLD/PSEL fields: read at a clock edge, so they only need to have settled by then. Crosstalk on them is mostly harmless.

The ranked list of what actually prevents trouble:

  1. Ground plane + row-B guard grounds (overwhelmingly the biggest factor)
  2. Pull-ups on the data bus — RN1 (8 × 10 k) in the schematic. Not a termination at all, but essential with HCT: DOE=0 leaves D0–D7 undriven, and floating CMOS inputs oscillate and burn power. 10 k parks idle lines at a legal high
  3. Clean clock distribution (guard traces; optionally a 33 Ω series resistor at the clock driver on the control card to soften the launched edge)

Why classic passive (Thevenin) termination is wrong for this bus: the 220/330 Ω pairs on S-100 and Multibus backplanes bias undriven lines to ~2 V — fine for LS-TTL inputs, but 2 V is dead in the middle of an HCT input's threshold region, recreating the floating-input problem on purpose, plus ~25 mA of standing current per terminated line. Don't copy it onto an HCT bus.

Active termination (a regulated ~2.7 V rail feeding ~120 Ω per line — the SCSI approach) fixes Thevenin's power waste but keeps the mid-threshold bias problem for CMOS and adds a regulator. Justifiable on a long, fast, heavily-loaded bus; overkill by an order of magnitude here. Done properly it is a whole signalling system rather than a part: low-swing open-drain transceivers (GTL/GTLP, BTL) with reference-voltage receivers. Bolted onto HCT it fixes nothing, and those transceivers are SMD-only, against the through-hole build.

Damped, not matched. Termination has to suit the driver's strength. An HCT bus driver (output resistance roughly 30–50 Ω) driving the 20–35 Ω loaded bus launches only about a third to a half of the swing; the reflection off the unterminated far end completes the edge. A matched far-end termination removes that reflection and turns a clean edge into a slow RC climb. So the provisioned networks are deliberately above the line impedance (100 Ω): they damp ringing without demanding matched-line current. The corollary: do not swap the bus drivers for a stronger, faster family (ACT, ABT) without revisiting this section — faster edges move the bus fully into the transmission-line regime, where it would then need real termination.

By signal class:

Class Lines Treatment
Clocks CLK, CLKB Spacing/guard on the board (§2); optional 22–33 Ω series at the driver (control card); 100 Ω + 150 pF at the far end (R2/C13, R3/C14), fitted only if the scope shows ringing
Async / strobes -RES, latch strobes Same as the clocks if a scope shows ringing
Open-drain, wired-OR IRQ The pull-up R4 (10 kΩ) is its termination. With all slots filled (≈ 100 pF) it rises in ≈ 1 µs; if release must be seen within a few cycles, drop R4 to 2.2 kΩ (≈ 0.2 µs, ≈ 2.3 mA per asserting driver)
Bidirectional data D0–D7 RN1 10 kΩ pull-ups (idle level, not termination). If ever needed, RC at both ends — any slot can drive
Address / fields A0–A15, DOE, DLD, PSEL None — sampled after settling

What the board provisions instead — AC (RC) termination on the two lines that merit it: CLK and CLKB each get 100 Ω + 150 pF to ground at the far end of the bus (R2/C13, R3/C14). AC termination damps reflections during edges but draws zero DC and imposes no bias level — the CMOS-friendly choice. Assembly guidance: leave them unpopulated at first. Bring the machine up, probe CLK at the far slot, and fit them only if you see ringing that crosses threshold regions. Footprints cost nothing; debugging mystery double-clocks costs weekends. If data/strobe lines ever ring problematically at higher clock speeds, the same RC treatment applies — but expect the planes to have already solved it.

4. Backplane BOM

Ref Part
J1–J8 DIN 41612 Type C 96-pin female, straight (vertical), THT solder — see the connector note below
J11 4-pos screw terminal, 5.08 mm
C11, C12 470 µF 10 V electrolytic
C1–C8 100 nF ceramic disc
RN1 10 kΩ × 8 bussed SIP-9
R4 10 kΩ ¼ W — IRQ pull-up (wired-OR, open-drain line)
R2, R3 100 Ω ¼ W (DNP initially)
C13, C14 150 pF ceramic (DNP initially)
R1, LED1 1 kΩ + 5 mm LED

The DIN 41612 connector pair

The backplane and the plug-in cards use a matched DIN 41612 Type C, 3×32 (96-way), 2.54 mm pitch, THT solder pair — one gender per side:

Side Footprint (KiCad Connector_DIN) Orientation Representative part (Harting)
Backplane (J1–J8) DIN41612_C_3x32_Female_Vertical_THT female, straight/vertical 09 03 296 6824 (96-pos female, 2.9 mm solder pins, THT)
Plug-in card (J1) DIN41612_C_3x32_Male_Horizontal_THT male, right-angle 09 03 196 6921 (96-way, Type C, 3-row, right-angle, solder)

Equivalent Type C 96-way solder parts are made by Amphenol (FCI), ERNI/TE and Assmann WSW, often better stocked at US distributors. Class 2 (tin tails) is adequate for a bench build; Class 3 adds gold plating and mating cycles.

Footprint desk-check done 2026-10-11: both footprints are the KiCad-stock Type C 3×32 pair — 2.54 mm column and row pitch, two M2.5 fixing holes (2.85 mm), 1.0 mm signal drill — and the pin numbering is the DIN 41612 standard on both, so a matched male/female pair mates a-a/b-b/c-c. Every card's J1 and every backplane slot assign pin X to busnet(X) from the one source, so a given pin carries the same signal on both sides. Before ordering the full set: confirm the exact manufacturer part number and live stock against its datasheet (body, fixing-hole pattern, THT-solder variant — not press-fit, not a contacts-not-included housing), and dry-fit one card male against one backplane female — which is why the backplane is ordered first.

Last updated 21 October 2026