Boards we signed off.

Three engagements from the last few quarters. Client names withheld at their request — the stats and outcomes are real.

14-layer · DDR5 6400 MT/s

High-speed digital

The DDR5 server that wouldn’t train.

14 layers
DDR5 6400 MT/s
28 diff pairs
±3 ps skew

ProblemA Series-B silicon startup’s vendor board topped out at DDR5-4800 and no one knew why. NPI was blocked on a thermal + SI sign-off.

What we foundStackup mistuned by 7 mil (Z₀ at 47 Ω vs. 50 Ω target); a BGA escape that added a via stub on the strobe; a hidden thermal choke under the DDR PHY.

What we didRe-tuned the stackup, back-drilled the offending vias, re-matched at the escape, added 12 thermal vias under U7, and shipped a full TRM thermal report.

Outcome: trained at full 6400 MT/s on first article. Eye height +38%. Thermal measured 79 °C vs. 82 °C predicted (≤4% error). NPI signed off 11 days later.

120 A · 3 oz Cu · EV charger

High power

120 A EV buck stage, re-spun before the melt.

6 layers
120 A continuous
3 oz copper
55 °C ambient

ProblemAn EV onboard charger’s buck stage ran hot in NPI. The team needed to know if 2 oz copper + heatsinks would survive a Phoenix summer.

What we foundTRM predicted 96 °C on the inductor pad at 2 oz / 55 °C — over the cap. Copper necked down at a routing channel; no thermal path for the high-side FET.

What we didRe-spec’d to 3 oz copper, widened the bottleneck 60%, added a copper coin under the FET, and re-ran the TRM sweep.

Outcome: hotspot dropped 96 → 79 °C in sim, measured 79 °C on bench. Saved a full NPI re-spin and a summer of field returns.

5.8 GHz · −110 dBm · mixed-signal

RF & microwave

5.8 GHz transceiver sharing a board with everything.

8 layers
Rogers 4350B
−110 dBm
50 Ω ±5%

ProblemA 5.8 GHz RX/TX had to share a board with a Cortex-M, a buck converter, and DDR — and still hit its noise floor.

What we foundThe only way to make budget was disciplined isolation: a Rogers sub-stack for the RF, a quiet reference island, and a switcher return path kept away from the LNA.

What we didHybrid Rogers/FR-4 build, stitched guard ring + shield-can footprint, EM-simulated connector launch matched to 50 Ω.

Outcome: measured noise floor −110 dBm — within 2 dB of the standalone reference design, on a board a quarter the cost of a separate RF module.

Names and identifying details withheld under NDA. We’ll walk you through the full reports under NDA on a call.

Your board could be next.

Senior EEs, TRM sign-off, first-article success. Fixed-fee band in 60 seconds.

Scope a project ↗See sign-off