
FujiNet NES Rev0
FujiNetWIFI/fujinet-hardware
Schematics and design files for FujiNet hardware

Companion KiCad project for "Fascia", a 10-inch Android Automotive touch display head unit: sibling to manifold-pcb, thermo-pcb and ecu-pcb, same 12V automotive constraints as the rest of the family.
Fascia is an open-source printed circuit board by jessiepullaro414, from the jessiepullaro414/Fascia repository on GitHub, published under the MIT licence. It is a 6-layer board of 147.8 × 71.1 mm with 118 parts on 129 nets, routed with 3,716 tracks and 247 vias.
It is built around the TLV767-Q1DRB. Its parts include 47 resistors, 38 capacitors, 9 ICs, 8 connectors and 4 transistors.
Companion KiCad project for "Fascia", a 10-inch Android Automotive touch
display head unit: sibling to
manifold-pcb,
thermo-pcb and ecu-pcb,
same 12V automotive constraints as the rest of the family.
Unlike the rest of the family this is a two-board project, because the panel is physically detached from the main board and sits across the dash from it:
ecu-pcb, line-level audio pre-outs to an external amplifier, a
microphone input, and USB-C for wired CarPlay / Android Auto.Bluetooth and Wi-Fi need no carrier hardware — they come with the
module's WB variant.
Module: Verdin iMX95 Hexa 8GB Wi-Fi / Bluetooth IT (Toradex PN 0089, 6x Cortex-A55, 8 GB RAM, 32 GB eMMC, -40 to +85 C, $627.25 at quantity 1-99 on 2026-09-30). Chosen over the 4 GB WB IT (PN 0226, $408.15) for headroom: Android Automotive plus a wired CarPlay / Android Auto projection stream is the heavy case, and RAM cannot be added later. The non-WB parts are out because the board relies on the module's wireless. The Verdin modules are pin-compatible, so this can be revisited without a board change.
PCB placed and routed to within 3 connections.
build_pcb.py generates Fascia.kicad_pcb: 118 footprints, 574 pads
netted across 129 nets, six layers, on a 148 x 71 mm board.
route_board.py (ported from the sibling boards) routes it with
FreeRouting, widens the power trunks, pours GND on In1 and In4, and then
asks kicad-cli DRC how many connections are really open. After a
150-pass run 3 connections are still unrouted: AOUT_RL_N (codec to
the audio connector), U9_ADR1_TIE, and LVDS_A_Y0N (bridge to panel
connector). FreeRouting oscillated between 3 and 4 for the last 100
passes, so more passes of the same run will not close them; they need a
second stochastic attempt or hand-routing. run_drc.py reports
schematic parity 0, 12 lib_footprint_mismatch findings that are a
KiCad tooling artifact, and one silkscreen-over-mask finding.
Nothing about the differential pairs is tuned: FreeRouting does not
length-match, so the DSI and dual-link LVDS pairs are ordinary routed
nets with no impedance control. route_board.py prints the resulting
intra-pair skew so it is a number, not a guess.
Placement uses the same skyline bottom-left packer as manifold-pcb
and thermo-pcb, ported rather than reinvented, with three additions
this board needed:
PIN_FIRST
in build_pcb.py), so the DSI-to-LVDS bridge sits beside the Verdin
socket instead of 75 mm away. Area order alone had put it there, and
nearly every open net in the first routing run was at that bridge.PULL). Once its anchor IC is
down, a passive takes the slot closest to it, with a height penalty so
it cannot strand a void below. Plain lowest-gap order had put a
pull-up 130 mm from the chip it belongs to. Estimated wirelength fell
from 7,612 to 5,895 mm (tools/placement_metric.py), and FreeRouting
at 12 passes went from 22 unrouted to 4.CLEAR) is 2.0 mm, up from 1.3. That alone took
12-pass routing from 22 unrouted to 11, at a cost of 7 mm of height.Group membership is derived from the netlist: each part joins whichever IC or connector it shares the most non-rail nets with. Rails are excluded deliberately: GND touches everything and would make every part look equally related to everything else.
The Verdin socket is rotated 90 degrees as part of that. Its footprint is 36 x 79 mm, tall and narrow; on its side it is 79 x 36 mm and the board becomes landscape, which is also the shape a dash opening wants.
Schematic complete — every block wired except the backlight.
build_schematic.py generates a real, KiCad-loadable schematic
containing the Verdin iMX95 X1 connector, the 12 V automotive front end,
the 5 V buck, the 1.8 V and 3.3 V rails, CAN FD, the SN65DSI85-Q1
DSI-to-LVDS bridge, the PCM3168A-Q1 audio subsystem with its level
shifters, the USB-C port, and the control/sequencing block. No PCB yet.
kicad-cli sch erc reports 10 violations, and every one is either
blocked on a datasheet that has not arrived or deliberately optional:
| What | Count | Why it is still open |
|---|---|---|
PWM_3_DSI, GPIO_10_DSI | 2 | Backlight — needs the iFan string voltage and current |
USB_2_* | 4 | Panel touch link — needs the panel's touch interface confirmed |
CAN_2_* | 2 | Optional second CAN bus; no consumer chosen |
DSI_IRQ | 1 | Bridge interrupt; wants a spare 1.8 V module GPIO |
IGN_SENSE | 1 | Ignition sense; wants a module ADC input |
160 pins carry real NoConnect items and 40 signal nets are wired.
Footprints are assigned and verified. Every part maps to a real footprint, and the build checks that each named file actually exists — a mistyped library path is otherwise invisible until the netlist reaches the PCB editor and silently drops the part.
TI's RJR0014A (VQFN-HR-14) has no KiCad library footprint, and it is not something a generic QFN generator can produce: the pads are all different sizes, unevenly spaced, and the four corner pads are L-shaped.
tools/build_lm61460_footprint.py builds it. The geometry is not
eyeballed off the drawing — the LAND PATTERN EXAMPLE on page 55 of TI
SNVSB70F is real vector art, so the pad rectangles were extracted from
the PDF's own path data and only the scale had to be inferred, from a
dimension the drawing states outright (pad 10 is 2.4 × 0.4 mm). That
scale then reproduced pad 10's 0.400 mm height exactly, which is what
makes the rest of the extraction trustworthy.
Each L-shaped pad is built from two rectangles sharing one pad number — a real KiCad construction this project has met before, in a Keystone fuse holder with two physical pads both numbered "1". It is far safer than approximating an L with its bounding box, which would put copper in the notch and eat the clearance to the neighbouring pads.
All four notches face the outer corner of the pattern, confirmed by rendering the drawing around pads 1, 9 and 11 rather than assuming the symmetry held.
The generator verifies itself twice: internally, that the pad table is
coherent (pin numbering, which pads are L-shaped, pad 10's dimensions,
overall extent, and that no two different pins overlap), and then by
loading the result back through KiCad's own engine. That second check
earned its place immediately — the first version emitted trailing
; PINNAME comments on the pad lines, and KiCad's s-expression format
has no comment syntax, so the whole library failed to load with nothing
more useful than "Unable to load library".
The Verdin module plugs into a standard DDR4 SODIMM socket, and KiCad's
SODIMM-260_DDR4_H4.0-5.2_OrientationStd_Socket has exactly 260 pads
numbered 1–260 — a direct match for the extracted pinout, so the biggest
footprint risk on this board turned out to be no risk at all.
Every one of these footprints numbers its exposed pad N+1 — 65 on the
TQFP-64s, 25 on the VQFN-24, 9 on the SON-8s. Each symbol therefore
carries an explicit EP pin, so what happens to that pad is a visible
decision on the schematic rather than an accident of the pad having no
net:
| Part | Exposed pad |
|---|---|
| LM74930-Q1 | NoConnect — must float. The datasheet forbids grounding it |
| TLV767-Q1, TPS2557-Q1 | GND |
| SN65DSI85-Q1, PCM3168A-Q1 | GND (PowerPAD is reference ground) |
The LM74930-Q1's NoConnect is load-bearing: without an explicit pin, that pad is merely netless, and nothing on the drawing would stop a later ground-stitching pass from claiming it.
What is already final:
NoConnect itemsA and
OUT on the shared source node, sense resistor with its 50 Ω tap into
CS+, OVCLAMP tied to OV for clamp-with-breaker rather than plain
disconnect, and the OV/UVLO dividers, timer, ILIM and IMON networksBIAS to the output rail (valid
because Vout ≤ 12 V), boot network through RBOOT, FB and EN dividers,
RT, and pull-ups on the open-drain FLT and PGOOD outputsSTB because the part's internal
pull-up would otherwise wake it in standbyCAN_1_TX / CAN_1_RX now reach the transceiverVCORE treated as the 1.1 V output it
is with its 1 µF, REFCLK resistor-pulled to ground since the pixel
clock comes from the D-PHY clock, ADDR strapped low, and the unused
DSI channel B inputs plus both RSVD pins left genuinely unconnected
with NoConnect items rather than tied offJ4 panel LVDS connector, provisional — carries all 20 LVDS
signals, but its real pin order and pin count await the panel datasheetJ5, microphone into the first differential ADC
input, TDM on DIN1, MODE strapped low for the I2C control portCTRL_PWR_EN_MOCI now gates the 1.8 V and 3.3 V LDOs, so
the carrier rails follow the module instead of racing it at power-on.Blocks are laid out independently, so two can drift into the same region and land one net's stub endpoint exactly on another's. That merges the two nets — electrically catastrophic and near-invisible on the drawing. It happened twice while building this schematic:
UVLO_DIV label;USB_ILIM stub landed on connector unit D's
USB1_VBUS_SENSE stub.ERC does report this as multiple_net_names, but only for the pair it
happens to notice and only after the fact. build_schematic.py now
records every terminated coordinate — pin positions and stub
endpoints, for both generic parts and the X1 connector — and fails
loudly, naming both nets and both owning references. The first version
of that check scanned only label positions and missed the second case
entirely, which is why it now tracks points rather than labels.
The PCM3168A-Q1 has two supply domains: analog VCC at 4.5–5.5 V and digital VDD at 3.0–3.6 V. VDD cannot be run at 1.8 V, so the codec cannot share the module's logic level. Reading the electrical table makes the consequence sharper than "it needs a translator":
| Codec | Verdin | |
|---|---|---|
| Input high, min | 2.0 V | drives 1.8 V max |
| Output high, min | 2.4 V | 2.1 V absolute maximum |
The first row means the module cannot reliably drive the codec. The second row is the dangerous one: the codec's output high is above the module's absolute maximum rating. Wired directly, the codec-to-module direction would not merely fail — it would damage the module.
So every digital line between them is translated: I2S and the clocks through SN74AXC4T245-Q1 translators, and I2C through the classic two-N-FET open-drain circuit, because a push-pull translator cannot pass a bus where either end may pull low.
What exists right now:
tools/extract_verdin_pinout.py — extracts the full 260-pin Verdin
iMX95 X1 connector pinout from Toradex's own hardware datasheet.verdin_pinout.py — the generated result: all 260 pins, 47 GND, 5 VCC,
and the 20 LVDS pins, cross-checked against a second independent table
in the same document.verdin_x1.py — functional banking of those 260 pins into multi-unit
symbol units, with a verifier that proves the banking is a true
partition of the connector.build_schematic.py — generates Fascia.kicad_sch, Fascia.kicad_sym,
sym-lib-table and (on first run only) Fascia.kicad_pro.Three things that were needed to get the first ERC pass down from 268 findings to the real ones, recorded because each cost a debug cycle:
pin_not_connected findings.
The fix is a short stub wire from the pin out to the symbol, which
also keeps the symbol graphic off the pin name.power_pin_not_driven (54 findings). Placing each flag on a real
point of its net clears it.kicad-cli needs a .kicad_pro to resolve ${KIPRJMOD}. Without
one it cannot find the project-local sym-lib-table and reports one
lib_symbol_issues warning per placed symbol (60 findings), even
though the symbols are embedded in the schematic and perfectly valid.Everything below the "Design decisions" section is a plan, not a description of something that exists. See "Known open items".
A 260-pin connector cannot be one schematic symbol — at 2.54 mm pitch a
single block is ~330 mm tall, taller than any sheet this project uses.
KiCad's answer is a multi-unit symbol, so verdin_x1.py assigns each pin
to a unit based on how this board uses the connector:
| Unit | Pins | Contents |
|---|---|---|
| A | 55 | Power and ground |
| B | 16 | Control, sequencing, JTAG, tamper |
| C | 15 | Display — the DSI link to the bridge, plus backlight control |
| D | 16 | Communications — CAN to ecu-pcb, I2C, USB 2.0 for touch |
| E | 158 | Unused on this board (gets NoConnect items) |
Banks are derived from pin names rather than hand-listed pin numbers,
so re-running the extractor against a newer datasheet revision cannot
silently desynchronise the two files. That only works if the rules stay
exhaustive, which verify() enforces: every pin exactly once, nothing
invented, nothing dropped, plus explicit assertions on the two groupings
most worth getting right —
USB_2_SS* and
module-specific USB1_TX1_*/USB1_RX1_* — and a single USB_ pattern
silently catches one and misses the other.Bank E is not filler. Per this project's KiCad notes a deliberately
unused pin needs a real NoConnect item, not a stub label, or ERC
reports it as a dangling/isolated label — so the unused pins have to be
enumerated as carefully as the used ones.
The project began as "10-inch Android Automotive display on an NXP i.MX 93". The i.MX 93 cannot run Android, and this was worth catching before any board work started:
Replaced with the i.MX 95, which has an Arm Mali-G310 (OpenGL ES 3.2, Vulkan 1.2, 2D+3D), is on NXP's Android BSP list, and is automotive qualified.
A raw i.MX 95 is a 19×19 mm BGA paired with LPDDR5. That forces HDI construction (laser microvias, blind/buried vias, 12–14 layers just to escape the BGA) and roughly 90 length- and skew-matched LPDDR5 nets.
That last part is the disqualifier for this project specifically: FreeRouting cannot do length-matched differential DDR routing, so the DDR work would have to be manual — which directly contradicts this family's rule that the generator scripts are the source of truth and the board is always regenerated, never hand-edited.
Putting the SoC on a Toradex Verdin iMX95 module moves the LPDDR5 and its HDI stackup onto a part Toradex has already qualified, and leaves a carrier board that is connectors, power and differential pairs — squarely inside what the existing generate → route → verify pipeline handles.
The panel is detached and 20 cm–1 m away across the dash, so the link that crosses that gap must be LVDS — differential and low-swing, so it survives alternator and ignition noise. Options for the long run:
| Option | Verdict |
|---|---|
| LVDS over shielded FFC | Chosen for the cable run. Robust over a dash-length ribbon, and what 10.1" panels natively accept. |
| MIPI DSI direct to the panel | Rejected. D-PHY is designed for on-PCB or very short flex (~15–20 cm); unshielded ribbon across a dash is an EMI problem. |
| HDMI | Not native on the module. Needs a DSI-to-HDMI bridge on the carrier and a receiver board behind the panel bridging back. Two extra bridges. |
| FPD-Link III / GMSL2 | The right answer for genuinely long runs (10 m+), overkill and expensive here. |
How the LVDS gets generated is the part that took real work.
The module's own display ceilings, from datasheet Table 3: LVDS 1× up to 1920×1080, MIPI DSI 1× quad lane up to 4K.
The 10.1" panel market is overwhelmingly 16:10 (1280×800 and 1920×1200), not 16:9. 1920×1200 needs ~154 MHz of pixel clock, which is above the module's 1920×1080 native LVDS ceiling — both Toradex and NXP's summary cap i.MX 95 LVDS at 1080p60. Forum figures of 170 MHz for dual-channel split mode are i.MX 6, a different generation, and must not be extrapolated to this part.
So: take MIPI DSI off the module (where 1920×1200 is far inside the 4K ceiling) and convert to dual-channel LVDS on the carrier board, right at the ribbon connector, with a TI SN65DSI85-Q1 — AEC-Q100 qualified, dual-channel LVDS out, 24 bpp, pixel clock 25–154 MHz.
The DSI run stays on-PCB where D-PHY is happy (centimetres); the long, noisy dash run is still LVDS. Cost is one bridge IC.
The module's native LVDS pins (88–144) are therefore unused. They
remain documented in verdin_pinout.py in case the design ever falls
back to 1080p native.
Unexpected benefit: the whole DSI pin group sits on Verdin-standard Reserved pins (17–55) rather than the Module-specific pins native LVDS uses. This design is therefore more portable to other Verdin modules than the native-LVDS version would have been — it inverts the portability caveat noted below.
Confirmed DSI pin group (Verdin iMX95 datasheet Table 22), which carries not just the lanes but every control signal the bridge and backlight need:
| X1 pin | Signal | Role |
|---|---|---|
| 23 / 25 | DSI_1_D3_N / _P | DSI data lane 3 |
| 29 / 31 | DSI_1_D2_N / _P | DSI data lane 2 |
| 35 / 37 | DSI_1_CLK_N / _P | DSI clock |
| 41 / 43 | DSI_1_D1_N / _P | DSI data lane 1 |
| 47 / 49 | DSI_1_D0_N / _P | DSI data lane 0 |
| 53 / 55 | I2C_2_DSI_SDA / SCL | SN65DSI85-Q1 configuration |
| 17 | GPIO_9_DSI | bridge enable |
| 19 | PWM_3_DSI | backlight brightness |
| 21 | GPIO_10_DSI | backlight enable |
The mode this whole design depends on is explicitly supported — datasheet Table 5, "Single DSI Input to Dual-Link LVDS":
Single DSI Input on Channel A to Dual-Link LVDS output with Odd pixels on Channel A and Even pixels on Channel B.
Register configuration for that mode:
| CSR | Field | Value |
|---|---|---|
| 0x10.6:5 | DSI_CHANNEL_MODE | 01 — single-channel DSI receiver |
| 0x10.4:3 | CHA_DSI_LANES | 00 — four lanes enabled |
| 0x18.4 | LVDS_LINK_CFG | 0 — dual-link LVDS |
| 0x18.3 | CHA_24BPP_MODE | 1 — 24 bpp |
| 0x18.1 / 0x18.0 | CHA/CHB_24BPP_FORMAT1 | selects Format 1 vs Format 2 |
Other confirmed facts that drive the schematic:
GPIO_9_DSI or by an RC (≈200 kΩ + ≈200 nF).
It must only go high after VCC is valid. To reset with VCC already
high, EN must be held low ≥10 ms.Y3 pairs are left NC; at 24 bpp all four data pairs
per link are used — so the panel connector carries 10 pairs.CHA_SYNC_DELAY must be programmed to at least 32 pixel clocks.Everything is configured over I2C — there is no format strapping.
That is a meaningful de-risking result: the VESA-versus-JEIDA mapping
question is a software choice (24BPP_FORMAT1), so picking the panel
later cannot force a board respin over it.
An earlier revision of this file claimed 1920×1200 sat "exactly at the bridge's 154 MHz ceiling". That was wrong, and the error mattered: the 25–154 MHz figure is the LVDS output clock per link, and in dual-link mode each link carries half the pixels. TI's own power-table reference case is literally 1920×1200 dual-link with LVDS CLK OUT = 81.6 MHz.
The real picture, with every number now confirmed against a primary source:
| Constraint | Value | Limit | Utilisation |
|---|---|---|---|
| i.MX 95 DSI transmit, per lane | ~925 Mbps | 2500 Mbps | ~37% — huge headroom |
| LVDS output clock, per link | ~81.6 MHz | 154 MHz | ~53% — comfortable |
| Bridge DSI input, per lane | ~925 Mbps | 1 Gbps | ~93% — the bottleneck |
The i.MX 95 side is confirmed a non-issue: IMX95AEC Rev 8, Table 58,
RATE[TX] Transmit Serial Data Rate = 80 to 2500 Mbps per lane, and
the feature summary lists "1x 350 MHz MIPI-DSI (4-lane, 2.5 Gbps/lane)
supporting 4kp30 or 3840 x 1440p60".
So the SN65DSI85-Q1 is the sole bottleneck in the chain. That is acceptable — TI documents 1920×1200 as a supported use case for this part — but it means the bridge, not the SoC, is what to re-examine if the timing ever fails to close.
The same datasheet confirms the native LVDS ceiling recorded above: "2x 1080p60 LVDS Tx (2x 4-lane or 1x 8-lane)".
VDD_LVDS_1P8-to-ground-via-10 kΩ guidance in
that same table applies to the SoC, which is Toradex's side of the
boundary, not this carrier's.Note the Verdin carrier must still follow Toradex's own Carrier Board Design Guide (doc 108140) for anything X1-side; UG10210 describes the raw SoC.
For reference, the module's native LVDS pins, should the fallback be needed:
| X1 pin | Channel 0 | X1 pin | Channel 1 |
|---|---|---|---|
| 88 / 90 | LVDS0_CLK_N / _P | 118 / 120 | LVDS1_CLK_N / _P |
| 94 / 96 | LVDS0_D0_N / _P | 124 / 126 | LVDS1_D0_N / _P |
| 100 / 102 | LVDS0_D1_N / _P | 130 / 132 | LVDS1_D1_N / _P |
| 106 / 108 | LVDS0_D2_N / _P | 136 / 138 | LVDS1_D2_N / _P |
| 112 / 114 | LVDS0_D3_N / _P | 142 / 144 | LVDS1_D3_N / _P |
Real parts found, with the specs that actually drive the board design:
Selected: iFan IF101GRL192-120B (2026-08-16), pending the vendor datasheet and pricing. Chosen mainly on temperature — it is a full 10 °C better at both ends than the other 1920×1200 candidate, which is the difference between reliable and marginal on a sun-baked dash — with 89% DCI-P3 gamut as a real bonus for image quality.
https://ifan-display.com/product/10-1-inch-touch-panel-1000-nits-wide-color-gamut-1920x1200/
| Part | Res | Nits | Interface | Touch | Bonded | Temp |
|---|---|---|---|---|---|---|
| iFan IF101GRL192-120B ← selected | 1920×1200 | 1000 | LVDS, 60-pin | PCAP inc. | ? | −30 to +80 °C |
| Riverdi RVT101HVLNWC00-B | 1280×800 | ~1000 | LVDS | PCAP inc. | yes | industrial |
| CDTech S101BWU78EP | 1920×1200 | 1000 | LVDS, 45-pin FPC | no | no | −20 to +70 °C |
| CDTech S101HWX101ED | 1280×800 | 1000 | 4-lane LVDS, 40-pin FPC | no | no | −30 to +80 °C |
| CDTech S101HWX80NP-FC81 | 1280×800 | 350 | 1-ch LVDS, 30-pin | ILI2511 (USB) | no (G+G) | −20 to +70 °C |
| AUO G101UAN02.0 | 1920×1200 | 800 | MIPI | no | no | −10 to +60 °C |
The market bifurcates, and neither half is the full target spec:
The full combination is realistically a semi-custom order with MOQ, not an off-the-shelf purchase.
AUO's G101UAN02.0 is ruled out on two counts despite being the obvious 1920×1200 name: it is MIPI, not LVDS (which would put D-PHY on the dash ribbon, the thing this whole design avoids), and −10 to +60 °C is unusable in a car.
Three things to carry forward:
The original plan assumed I2C touch. The market says otherwise, and on reflection USB is probably the better engineering choice here anyway:
Decision still open, but it should be made on cable length, not on which is simpler on paper.
Every part below was verified against its own datasheet, not chosen from memory. Two of them corrected an earlier wrong guess, noted inline.
| Role | Part | Key evidence |
|---|---|---|
| Front-end protection | LM74930-Q1 | AEC-Q100 grade 1 (−40/+125 °C), 4–65 V in, reverse protection to −65 V, adjustable OC/OV, circuit breaker, "meets automotive ISO7637 transient requirements with a suitable TVS diode" |
| 12 V → 5 V buck | LM61460-Q1 | AEC-Q100 grade 1, 3–36 V in (42 V transient), 6 A, synchronous, low-EMI with spread spectrum |
| CAN FD | TCAN1044V-Q1 | AEC-Q100 grade 1, VIO 1.7–5.5 V with explicit 1.8 V support, CAN FD to 8 Mbps, ±58 V bus fault protection |
| DSI→LVDS bridge | SN65DSI85-Q1 | AEC-Q100 grade 2 (−40/+105 °C), single 1.8 V rail, HTQFP-64 |
| Audio codec | PCM3168A-Q1 | AEC-Q100, 24-bit 6-in/8-out codec, differential outputs, I2S/TDM, 104 dB, 64-HTQFP |
| USB VBUS switch | TPS2557-Q1 | Automotive, current limit adjustable 500 mA–5 A, FAULT output, dual thermal sensing |
| Wi-Fi / Bluetooth | on-module (u-blox MAYA-W260-00B) | Wi-Fi 6, Bluetooth Classic/LE 5.4, two U.FL connectors — comes with the Verdin WB variants |
The module sets two hard constraints that drove these choices:
TI's own front-page reference circuit for the LM74930-Q1 is titled "VBAT 12-V or 24-V With 200-V Unsuppressed Load Dump — Output Clamp". That is this application almost exactly: an unsuppressed alternator on a 1972 car, with the output clamped rather than disconnected.
The back-to-back FET arrangement is easy to get backwards, so it is written down here:
VBAT ──[RSENSE]──┬── Q2 drain Q2 = pass / load switch (HGATE)
│ Q1 = ideal diode (DGATE)
Q2 source ──┬── COMMON ──┬── Q1 source
│ │
OUT (15) A (2) ← both pins on COMMON
(common-source pair)
Q1 drain ── VOUT_PROT ── C (24)The resolution: A (2) and OUT (15) are the same node. "Back-to-back
N-channel MOSFETs in common source configuration" means both FET sources
tie together, and the datasheet independently says A is "the source of
the external ideal diode MOSFET" while OUT is "the common source rail".
C (24) is the ideal diode's drain and therefore the protected output.
Direction is confirmed by the ideal-diode spec regulating "10.5 mV A to
C forward voltage drop", and by HGATE driving "the first MOSFET in
the power path".
Reference-design support components, all named in that circuit: RSENSE
with RSET (50 Ω into CS+), RISCP on ISCP, RILIM on ILIM,
RMON on IMON, CT on TMR, CVS on VS, CCAP across CAP–VS,
and the R1/R2/R3 dividers feeding OV/OVCLAMP and UVLO.
Rough budget, with the module figure still a placeholder:
| Load | Estimate |
|---|---|
| Verdin iMX95 | ~3 A @ 5 V (unknown — see open items) |
| USB-C charging | 3 A @ 5 V |
| Bridge, codec, CAN, misc | ~0.5 A |
| Total on 5 V | ~6.5 A |
That is over the LM61460-Q1's 6 A, so two supplies rather than one bigger one: a main rail for the module and peripherals, and a second dedicated to USB VBUS. Splitting is the better design anyway — a phone drawing a fault current should not be able to brown out the SoC and reboot Android. Using the same part twice also keeps the BOM narrow.
The backlight boost is deliberately not on the 5 V rail — it runs from the protected 12 V directly, since the panel wants ~27–34 V and routing that through 5 V would waste a conversion and eat the budget.
The Verdin WB variants carry a u-blox MAYA-W260-00B: Wi-Fi 6,
Bluetooth Classic / LE 5.4, IEEE 802.15.4, and two U.FL antenna
connectors already on the SoM. That means no RF design, no antenna
matching and no radio certification work on this carrier — just two U.FL
cables out to external antennas.
Bluetooth Classic covers A2DP for streaming and HFP for hands-free calls. Wi-Fi plus Bluetooth together is also exactly what wireless CarPlay and Android Auto need, so the USB-C port is for wired mode and charging rather than being the only way in.
Cost to record: the datasheet carries separate pin tables for "modules
without Wi-Fi", so the WB variant consumes some UART and ADC pins that
would otherwise reach the carrier. The variant has to be locked before
the connector banks are final.
The CAN transceiver must be the V variant. An earlier draft of this
plan named the TCAN1043-Q1, which only level-shifts to 3.3 V or 5 V — it
cannot talk to 1.8 V logic at all. Worse, within the correct family the
datasheet's own device comparison table reads:
| Part | Low-voltage I/O support on pin 5 |
|---|---|
| TCAN1044-Q1 | No — pin 5 is a No-Connect |
| TCAN1044V-Q1 | Yes — pin 5 is VIO, 1.8–5.5 V |
The suffix is the entire difference. Order the wrong one and the board assembles perfectly and the CAN link never works.
The surge stopper is TI's, not the obvious ADI part. The LTC4364 is
the better-known automotive surge stopper and does something the
LM74930-Q1 does not — it regulates through an overvoltage event rather
than disconnecting. But its datasheet order table lists only C, I and H
temperature grades with no AEC-Q100 qualification, which fails this
family's standard (see ecu-pcb, where the last unqualified part was
deliberately swapped out). The LM74930-Q1 is explicitly -Q1.
Correction — the LM74930-Q1 does clamp through. An earlier revision of this file claimed it protects only by disconnecting the load, and warned that a load dump would therefore collapse the 5 V rail and force an Android reboot. That was wrong, and reading the pin table and spec table settled it:
OVCLAMP: "connect this pin to OV pin for overvoltage
clamp with circuit breaker (timer) functionality";OVCLAMP has its own rising/falling thresholds (0.59 V / 0.45 V),
separate from the plain OV disconnect thresholds;I(TMR_SRC_OVCLAMP), "TMR source current
during overvoltage clamp" — a dedicated clamp state, not a
disconnect;N(A_R_Count) gives 32 auto-retry cycles.So it regulates the pass FET through an overvoltage event for a period set by the TMR capacitor, then opens the breaker if the event outlasts it, then auto-retries. That is genuine surge-stopper behaviour, and it removes the reboot risk previously recorded here.
What remains real is that the clamp is timed, not indefinite: the TMR capacitor has to be sized to ride out a 200 ms load dump, and the pass FET has to survive dissipating it for that long. FET SOA selection is therefore a genuine design task — see "Known open items".
For scale, the LTC4364 datasheet gives the industry-standard transient shapes: a general automotive transient of tr = 10 µs / VPK = 80 V / τ = 1 ms, and load dump of tr = 5 ms / VPK = 60 V / τ = 200 ms. The 200 ms one carries the energy that matters.
Recorded because it should drive the panel budget. In a car, in rough order of impact on how good the display actually looks:
Resolution ranks below all of these. The target spec is a 1080p, ~1000-nit, optically-bonded, 24-bit IPS panel.
Caveat worth recording: LVDS on Verdin is a Module-specific
function, not an Always-Compatible or Reserved one (datasheet Table 6
classifies it Total 1 / Always Compatible 0 / Reserved 0 / Module-specific 1). The pins are genuinely on the connector, but this
carrier is therefore tied to the Verdin iMX95 specifically and is not
portable to an arbitrary other Verdin module. MIPI DSI, by contrast, sits
on Verdin-standard Reserved pins (17–55) and would be portable — at the
cost of needing a bridge for a dash-length run.
This is also why tools/extract_verdin_pinout.py exists rather than a
hand-typed table: the LVDS pins are named MSP_1…MSP_25 in the Verdin
specification column and only resolve to LVDS0_CLK_N etc. in the
module-specific function column. A parse that reads the wrong column
concludes the module has no LVDS at all.
VCC_BACKUP (249).CAN_1_TX/CAN_1_RX on pins 20/22, CAN_2 on 24/26.I2C_2_DSI on 53/55.PWM_3_DSI (pin 19) for brightness, GPIO_10_DSI (pin 21)
for backlight enable via the on-module PCAL6416A I/O expander.A head unit that displays and sources audio but does not amplify it. Scope revised 2026-08-16 from the original display+touch+CAN.
In scope: display, touch, CAN to ecu-pcb, Bluetooth, line-level
pre-outs to an external amplifier, a microphone for calls and voice
control, and USB-C for wired CarPlay / Android Auto plus phone charging.
Explicitly not in scope: any power amplifier. This board drives no speakers — it feeds an existing automotive amp over RCA-level pre-outs. That single decision keeps the highest-current, highest-heat, highest-EMI subsystem off the board entirely, which matters in a sealed dash enclosure that already contains a 1000-nit backlight.
Also still out of scope: reversing camera, automotive Ethernet, and any radio tuner.
Carrier board blocks:
CTRL_PWR_EN_MOCI, CTRL_RESET_MOCI#,
CTRL_SLEEP_MOCI#, ignition/ACC sense.I2C_2_DSI, enabled by GPIO_9_DSI.USB_1 port, and VBUS from a TPS2557-Q1 switch whose EN
and FAULT map onto USB_1_EN and USB_1_OC#. CC1/CC2 carry Rp
resistors advertising 3 A as a downstream-facing port — no PD
controller and no negotiation firmware.PWM_3_DSI, enabled by
GPIO_10_DSI, sized for a high-brightness (~1000 nit) panel rather
than a consumer-grade one.ecu-pcb.Panel board blocks:
Same script-driven, kiutils/kicad-cli-verified workflow as the sibling
projects: the generator scripts are the source of truth — always
regenerate from them, never hand-edit Fascia.kicad_sch/Fascia.kicad_pcb
directly.
tools/extract_verdin_pinout.py is a one-shot data extractor, not part of
the per-build pipeline. It takes the Toradex datasheet's local path as an
argument (the PDF is Toradex's document and is gitignored, not committed):
python tools/extract_verdin_pinout.py ~/Downloads/200007-verdin_imx95_datasheet.pdfIt refuses to write output unless all 260 pins parse and the LVDS pin numbers it derived from the main pin-assignment tables agree with the separate LVDS table on page 42.
| Qty | Value | Footprint | References |
|---|---|---|---|
| 6 | 100n 1V8 | Capacitor_SMD:C_0603_1608Metric | C16, C17, C18, C19, C20, C21 |
| 4 | 100n codec | Capacitor_SMD:C_0603_1608Metric | C31, C32, C33, C34 |
| 4 | 10k pullup | Resistor_SMD:R_0603_1608Metric | R44, R45, R46, R47 |
| 4 | 10k tie-off | Resistor_SMD:R_0603_1608Metric | R33, R34, R35, R36 |
| 4 | 2k2 pullup | Resistor_SMD:R_0603_1608Metric | R25, R26, R27, R28 |
| 3 | 10u decouple | Capacitor_SMD:C_0603_1608Metric | C28, C29, C30 |
| 3 | 10u in | Capacitor_SMD:C_0603_1608Metric | C5, C9, C22 |
| 3 | FB bot | Resistor_SMD:R_0603_1608Metric | R11, R18, R24 |
| 3 | FB top | Resistor_SMD:R_0603_1608Metric | R10, R17, R23 |
| 2 | 100k FLT pu | Resistor_SMD:R_0603_1608Metric | R15, R38 |
| 2 | 10k bias | Resistor_SMD:R_0603_1608Metric | R31, R32 |
| 2 | 10u out | Capacitor_SMD:C_0603_1608Metric | C10, C23 |
| 2 | 60R term | Resistor_SMD:R_0603_1608Metric | R19, R20 |
| 2 | NFET i2c xlat | Package_TO_SOT_SMD:SOT-23 | Q3, Q4 |
| 2 | SN74AXC4T245-Q1 | Package_SO:TSSOP-16_4.4x5mm_P0.65mm | U7, U8 |
| 1 | 0R ID | Resistor_SMD:R_0603_1608Metric | R41 |
| 1 | 100k PG pu | Resistor_SMD:R_0603_1608Metric | R16 |
| 1 | 100k RST | Resistor_SMD:R_0603_1608Metric | R29 |
| 1 | 100k Vsense | Resistor_SMD:R_0603_1608Metric | R42 |
| 1 | 100n | Capacitor_SMD:C_0603_1608Metric | C39 |
| 1 | 100n CCAP | Capacitor_SMD:C_0603_1608Metric | C3 |
| 1 | 100n CVS | Capacitor_SMD:C_0603_1608Metric | C2 |
| 1 | 100n IN | Capacitor_SMD:C_0603_1608Metric | C37 |
| 1 | 100n RST | Capacitor_SMD:C_0603_1608Metric | C35 |
| 1 | 100n U7A | Capacitor_SMD:C_0603_1608Metric | C24 |
| 1 | 100n U7B | Capacitor_SMD:C_0603_1608Metric | C25 |
| 1 | 100n U8A | Capacitor_SMD:C_0603_1608Metric | C26 |
| 1 | 100n U8B | Capacitor_SMD:C_0603_1608Metric | C27 |
| 1 | 100n VCC | Capacitor_SMD:C_0603_1608Metric | C11 |
| 1 | 100n VIO | Capacitor_SMD:C_0603_1608Metric | C12 |
| 1 | 100n boot | Capacitor_SMD:C_0603_1608Metric | C7 |
| 1 | 100uF | Capacitor_SMD:C_0603_1608Metric | C1 |
| 1 | 10k JTAG Vref | Resistor_SMD:R_0603_1608Metric | R48 |
| 1 | 10k REFCLK | Resistor_SMD:R_0603_1608Metric | R22 |
| 1 | 10k Rp CC1 | Resistor_SMD:R_0603_1608Metric | R39 |
| 1 | 10k Rp CC2 | Resistor_SMD:R_0603_1608Metric | R40 |
| 1 | 10k STB pd | Resistor_SMD:R_0603_1608Metric | R21 |
| 1 | 10u 1V8 | Capacitor_SMD:C_0603_1608Metric | C15 |
| 1 | 150u VBUS | Capacitor_SMD:C_0603_1608Metric | C38 |
| 1 | 1u VCC | Capacitor_SMD:C_0603_1608Metric | C6 |
| 1 | 1u VCORE | Capacitor_SMD:C_0603_1608Metric | C14 |
| 1 | 1u couple | Capacitor_SMD:C_0603_1608Metric | C36 |
| 1 | 2.2u | Inductor_SMD:L_1210_3225Metric | L1 |
| 1 | 2k2 mic bias | Resistor_SMD:R_0603_1608Metric | R30 |
| 1 | 2m sense | Resistor_SMD:R_0603_1608Metric | R1 |
| 1 | 36k ILIM | Resistor_SMD:R_0603_1608Metric | R37 |
| 1 | 44u out | Capacitor_SMD:C_0603_1608Metric | C8 |
| 1 | 47k min | Resistor_SMD:R_0603_1608Metric | R49 |
| 1 | 4n7 split | Capacitor_SMD:C_0603_1608Metric | C13 |
| 1 | 50R | Resistor_SMD:R_0603_1608Metric | R2 |
| 1 | 56k Vsense | Resistor_SMD:R_0603_1608Metric | R43 |
| 1 | 5A | Fuse:Fuse_Bourns_MF-RG300 | F1 |
| 1 | 5k RMON | Resistor_SMD:R_0603_1608Metric | R4 |
| 1 | Buttons | Connector_PinHeader_2.54mm:PinHeader_1x04_P2.54mm_Vertical | J9 |
| 1 | CAN to ECU | Connector_PinHeader_2.54mm:PinHeader_1x03_P2.54mm_Vertical | J3 |
| 1 | EN bot | Resistor_SMD:R_0603_1608Metric | R14 |
| 1 | EN top | Resistor_SMD:R_0603_1608Metric | R13 |
| 1 | ESD DN | Diode_SMD:D_SMB | D3 |
| 1 | ESD DP | Diode_SMD:D_SMB | D2 |
| 1 | ESD VBUS | Diode_SMD:D_SMB | D4 |
| 1 | JTAG debug | Connector_PinHeader_2.54mm:PinHeader_1x08_P2.54mm_Vertical | J8 |
| 1 | LM61460-Q1 | TI_RJR0014A_VQFN-HR:TI_RJR0014A_VQFN-HR-14_4x3.5mm | U2 |
| 1 | LM74930-Q1 | Package_DFN_QFN:Texas_RGE0024H_VQFN-24-1EP_4x4mm_P0.5mm_EP2.7x2.7mm | U1 |
| 1 | Mic | Connector_PinHeader_2.54mm:PinHeader_1x03_P2.54mm_Vertical | J6 |
| 1 | NFET diode | Package_TO_SOT_SMD:SOT-23 | Q1 |
| 1 | NFET pass | Package_TO_SOT_SMD:SOT-23 | Q2 |
| 1 | OV bot | Resistor_SMD:R_0603_1608Metric | R6 |
| 1 | OV top | Resistor_SMD:R_0603_1608Metric | R5 |
| 1 | PCM3168A-Q1 | Package_QFP:Texas_TQFP-64-1EP_10x10mm_P0.5mm_EP8x8mm_Mask4.44x4.44mm | U9 |
| 1 | Panel LVDS (provisional) | Connector_PinHeader_2.54mm:PinHeader_1x22_P2.54mm_Vertical | J4 |
| 1 | Power in | Connector_PinHeader_2.54mm:PinHeader_1x03_P2.54mm_Vertical | J2 |
| 1 | Pre-outs to amp | Connector_PinHeader_2.54mm:PinHeader_1x11_P2.54mm_Vertical | J5 |
| 1 | RBOOT | Resistor_SMD:R_0603_1608Metric | R9 |
| 1 | RILIM | Resistor_SMD:R_0603_1608Metric | R3 |
| 1 | RT | Resistor_SMD:R_0603_1608Metric | R12 |
| 1 | RTC coin cell | Battery:BatteryHolder_Keystone_1058_1x2032 | J10 |
| 1 | SN65DSI85-Q1 | Package_QFP:Texas_TQFP-64-1EP_10x10mm_P0.5mm_EP8x8mm_Mask4.44x4.44mm | U5 |
| 1 | TCAN1044V-Q1 | Package_SO:SOIC-8_3.9x4.9mm_P1.27mm | U4 |
| 1 | TLV767-Q1 | Package_DFN_QFN:Texas_DRB0008A | U3 |
| 1 | TLV767-Q1 3V3 | Package_DFN_QFN:Texas_DRB0008A | U6 |
The 111 most common of 118 parts.

FujiNetWIFI/fujinet-hardware
Schematics and design files for FujiNet hardware


Josephrp/meshtonic
Meshtonic H4M Companion — ESP32-S3 carrier for PortaPack H4M / Mayhem with multi-slot LoRa and standalone Meshtastic


nathanfraske/CEC-Platform
ONE board file, kept current by the wave (owner directive 2026-07-25: "the current best should be placed into a candidate folder per board and kept current with only one board ideally so we have a reference").



It measures 147.8 × 71.1 mm and has 6 copper layers, 1.6 mm thick.
118 parts, built around the TLV767-Q1DRB. The full list of its parts is on this page, with links to each part's symbol and footprint where PCBKit's library has them.
From the jessiepullaro414/Fascia repository on GitHub, at the commit this page was made from: the board is Fascia.kicad_pcb and its schematic Fascia.kicad_sch. PCBKit opens the board in the browser, with nothing to install.
Its repository publishes it under the MIT licence: you may reuse it on that licence's terms (keep the author's notice, and for a copyleft licence share your changes alike).