← Insights

Case

Review of CERN's DI/OT Zynq UltraScale+ System Board

An independent, published review of a 14-layer FPGA carrier board — 67 findings, every one cited. The kind of thing ERC and DRC don't see.

2026-11-19 · 9 min · Case

All findings are taken from the published review at cern.probe.esharp.se.

We had Probe read one of the most advanced boards in CERN’s open hardware catalogue: the DI/OT Zynq UltraScale+ System Board (OHWR0083, revision R5). A 14-layer carrier board of 222.4 × 100 mm with 1,104 components and 1,208 nets, drawn in Altium Designer and published under CERN-OHL-W-2.0. It isn’t our own board — and that’s the point. The review is independent, and it’s open to browse.

Probe reads a board the way a senior engineer would: the layout, the schematic and each component’s datasheet together. It checks what was built against what each part requires, and shows each finding with the evidence behind it. On this board the tally came to 67 findings — 1 blocker, 43 major and 23 minor. Here are six of them, and why they’re hard to catch with rule checks.

A blocker the netlist hides

The net GND ties together a supply pin (IC30 pin 22) and a ground pin (IC48 pin 3) — a direct short between supply and ground. An ERC sees two pins both legally allowed to be called GND and says nothing. Probe reads what the pins are in their respective datasheets, sees that one supplies and the other grounds, and raises it as a blocker with both pins pointed out.

Reset released before the supplies are up

This is the finding that shows what “reading the schematic and the datasheet together” means.

PS_POR_B — the power-on reset to the Zynq device’s processing system — is released about 400 ms after P3V3 comes up. But the datasheet (DS925, Table 37) is clear: PS_POR_B must be held low until all PS supplies have reached their minimum level. On the board, the power-good outputs for P0V85, P1V2, P1V8 and VCC_PSPLL from the voltage regulators IC30 and IC31 are simply not wired in. If one of those supplies sags halfway — or comes up late — the Zynq is released from reset anyway, partly powered.

Probe traces the whole power-good chain pin by pin, compares it with the datasheet requirement, and also notes that the problem is known: it matches CERN’s own issue #385 and persists in V5 (it was already there in V4). It’s not something a geometric rule can find, because nothing in the layout is “wrong” — the fault lies in what the connectivity means against what the device requires.

When the protection becomes the obstacle

The USB-JTAG translator IC4 (SN74AVC2T245) drives PS_SRST_B push-pull through R324 (0 Ω), straight at the supervisor IC27 (MAX16025) which drives the same net open-drain. With a USB cable plugged in and no JTAG pod in place, the push-pull-driven high wins over the supervisor’s open-drain low: both the 800 ms power-on hold and the manual PB3 reset are defeated, and IC27’s output is driven to its 20 mA limit. Seeing it requires reading the drive strength in both datasheets and understanding who wins — not two nets being allowed to touch.

Noise that leaves the board

Probe lists 42 cases of off-board coupling. One example: ETH_TXD3 and FMC.LA_N25 run over each other on layers L7 and L8, and the Ethernet edge puts estimated noise onto the LVDS net. That net has a pin on the FMC connector — so the noise follows off the board, onto whatever is seated in the connector. (The magnitudes rest on Probe’s assumed edge rates, not on sourced values, and are to be confirmed; the coupling itself is read from the copper.) A crosstalk rule looks within the board; this finding is about what leaves it.

Components that have already gone end-of-life

A review that reads the datasheet also reads the lifecycle. On this board:

The I²C mux IC6 (PCA9548APW, NXP) is reported obsolete — last order date passed back on 2018-01-24, none left at the distributor and no approved replacement registered. The DDR4 memory MT40A1G16 sits in six placements and is obsolete. The configuration flash MT25QU01G likewise. And a TVS diode, BZX84C3V3, went from “Active” to “Obsolete” while the board was in review — a procurement risk that arose after the design was frozen. None of this is a layout or schematic question; it’s a question of what can actually be bought.

Why ERC and DRC don’t see this

The six findings have one thing in common: they require reading layout, schematic and datasheet together, and weighing in intent, change history and the components’ lifecycle.

An ERC checks that the netlist is consistent. A DRC checks that the geometry holds the manufacturing rules. Both are necessary, and neither knows what a pin is for, what a datasheet requires, which fix an issue trail asked for, or whether a device is still buyable. That’s where a review like Probe’s picks up. More on that boundary in What a review finds that DRC misses.

Open, for anyone who wants to browse

The whole review of the DI/OT SB ZU — and of 87 other CERN designs, together 1,351 findings — is public, each finding tied to the pins, nets and datasheet pages it rests on:

cern.probe.esharp.se →

The designs are owned by their authors and published on the CERN Open Hardware Repository under CERN-OHL. Probe is run by E# (E-Sharp AB).

Want to see the same on your own design? You run the review yourself — book a demo.

Book a demo of Probe

Show us your board — we review it and go through the findings together.

Prefer to email? info@esharp.se