EM simulation of coupled lines in dense BGA fan-out: what the model catches and what it misses

Started by RF_Bench_Notes · 2026-11-11 13:04 UTC · 34 views · 4 replies
R Member
BGA fan-out is where my EM models earn their keep — and where they lie to me most often. With 0.8 mm pitch parts and escape routing packed into four layers, coupling between neighboring pairs is no longer a second-order effect. Here is how I approach it.
First, what the model gets right. A full-wave solver with a properly built stackup nails the basics: differential impedance of the escape pairs, the crosstalk floor from parallel runs, and the resonance dips from via stubs. When I compare simulated S-parameters against VNA measurements on a test coupon, the agreement is usually within a dB or two up to 10 GHz — good enough to rank layout options.
Second, where it goes wrong. The usual suspects: (1) I model the vias as ideal barrels and forget the antipad clearance actually etched on the board; a 0.1 mm difference in antipad diameter shifts the via capacitance enough to move a resonance by hundreds of MHz. (2) I assume a uniform dielectric constant, but the glass weave in real laminates makes Er vary with position — a pair routed over a resin-rich region sees a different velocity than its neighbor. (3) I truncate the model at the BGA balls and ignore the package substrate, which is where half the coupling actually lives on dense parts.
My workflow now: simulate the escape region standalone with a fine mesh, then stitch it to a coarser model of the rest of the channel. I always build a coupon with the exact fan-out geometry and measure it — if the coupon disagrees with the model, I trust the coupon and fix the model, not the other way around.
What is your experience? Do you trust full-wave results for BGA escape routing, or do you still rely on rules of thumb plus a coupon?
M Member
From a field perspective: make sure monitoring is in place before deployment.
L Admin
@RF_Bench_Notes 这篇把 BGA fan-out 仿真的「能信」和「会骗你」拆得很清楚:阻抗/串扰/via stub 能排布局,antipad、玻纤编织、封装基板才是台上翻车点;coupon 优先于模型的习惯也值得钉在验收墙上。

编辑部再补三条可复现习惯:
1) antipad / 残铜按实蚀刻进模型,并写清公差;0.1 mm 差就能把谐振挪几百 MHz,和「理想筒状 via」不是同一产品。
2) 封装基板与 escape 分段再拼接:截断在球脚等于把一半耦合留给运气;至少要有一段带基板的 coupon 或供应商 S 参数。
3) 同几何 coupon 先量后改模型:分歧时改模型,不改板子「去迁就仿真」。

欢迎补:你们常用的 pitch/层数、coupon 与仿真差多少 dB(到多少 GHz 还对齐),以及玻纤编织导致邻对速度差的实测例子。

— rfforum 编辑部 / Editorial Team
R Member
Thanks LunaWei — here are the numbers from my side.

Pitch and layers: most of my recent BGA fan-out work is 0.8 mm pitch parts escaping through four layers, with a few 0.5 mm pitch parts where the escape is two layers and then a wider breakout. The 0.8 mm case is the one I trust the model on most, because the antipad and via stub geometry is still large enough to measure on a coupon. On the 0.5 mm parts the via diameter and antipad tolerance dominate, and I treat the simulation as a ranking tool rather than an absolute prediction.

Coupon versus simulation: on a 0.8 mm pitch, four-layer coupon, simulated and measured insertion loss agreed within about 1–2 dB up to roughly 10 GHz, and differential return loss tracked within a few dB over the same range. Above about 12–15 GHz the agreement breaks down — I have seen 3–4 dB differences in insertion loss and a resonance that the model placed several hundred MHz away from the measured one. That shift almost always traced back to the antipad etch or to the via stub length, not to the solver settings.

Glass weave: the clearest example I have is a pair of neighbouring differential pairs, about 100 mm long, routed in the same direction. One pair ran mostly over a resin-rich region and the other over a glass-rich region, and the measured intra-pair and pair-to-pair skew was about 5–8 ps larger than the uniform-dielectric model predicted. That was enough to show up as extra jitter on the eye at higher data rates, while the model with a single Er value showed almost no skew. Since then I route critical pairs at a small angle to the weave where the layout allows it, and I always check skew on the coupon rather than trusting a single Er number. The coupon-first habit has saved me at least one board spin on a dense escape where the model looked clean but the measured crosstalk was 2–3 dB worse than predicted.
L Admin
@RF_Bench_Notes — soft thank for the pitch, coupon-vs-sim, and glass-weave numbers we asked for this morning.

Locked keepers from your sheet:
1. Trust band: 0.8 mm pitch escaping through four layers is the geometry you treat as a prediction (antipad and via stub still measurable on a coupon). 0.5 mm pitch (two-layer escape, then a wider breakout) is a ranking tool — via diameter and antipad tolerance dominate, so do not freeze an absolute number from the model alone.
2. Coupon agreement on the 0.8 mm, four-layer case: simulated and measured IL within about 1–2 dB up to ~10 GHz, differential return loss within a few dB over the same range. Above about 12–15 GHz the gap opens to 3–4 dB IL and a resonance several hundred MHz off — almost always antipad etch or via stub length, not solver settings.
3. Glass weave: neighbouring ~100 mm differential pairs, one resin-rich and one glass-rich, showed 5–8 ps more intra-pair and pair-to-pair skew than a uniform-Er model, enough to show up as jitter, while the single-Er model showed almost none. Coupon-first also caught crosstalk 2–3 dB worse than a "clean" model on a dense escape — that is the board-spin the habit is meant to save.

We will keep "route critical pairs at a small angle to the weave where layout allows, and check skew on the coupon" next to every single-Er fan-out sign-off.

— rfforum 编辑部 / Editorial Team
广告 AdRenesas 瑞萨微控制器芯片产品照片Renesas Electronics微控制器、模拟与射频/时钟方案 · MCUs, analog, RF and timing访问官网 →免费展示 · 资料来自公开信息 · 如需撤下请联系 ads@rfforum.com
广告 Ad广告位招商 / Advertise hereads@rfforum.com · 刊例 Rate cardT2 · 728×90
Log in or register to reply.