科研 Skill

高速信号完整性分析

signal-integrity-analysis

Analyzes high-speed PCB signal integrity: impedance stack-up, S-parameter extraction, eye-diagram budgeting, and crosstalk checks. Use when routing USB3/HDMI/DDR/SerDes links.

发布者 Codar安全审核通过v0.1.0更新于 2026-11-18

SKILL.md

Signal Integrity Analysis

Verifies that a high-speed digital or mixed-signal PCB link will actually work at speed, not just that it is DRC-clean. Signal integrity (SI) failures (reflections, crosstalk, excessive jitter, ground bounce) routinely pass schematic review and DRC yet still fail bring-up or compliance testing. This skill gives the concrete analysis steps, tools, and pass/fail numbers to catch those failures on paper before fabrication.

When to use this

  • Designing or reviewing any interface above roughly 100 Mbps per lane: USB 2.0 Hi-Speed/3.x, HDMI, DisplayPort, PCIe, SATA, Ethernet (RGMII/SGMII/ 10GBASE-KR), DDR3/DDR4/LPDDR, MIPI (CSI/DSI), or any SerDes.
  • A board bring-up shows link training failures, bit errors, CRC errors, or an eye diagram that does not open, despite power rails and schematic being correct.
  • Stack-up or routing decisions (trace width, layer assignment, via count, connector choice) must be made before PCB release and need a quantitative basis instead of a rule of thumb.
  • A compliance lab pre-check is wanted before paying for formal certification testing (e.g. USB-IF, HDMI compliance).

This is a design-verification method, not a PCB layout tool: pair it with pcb-engineer or pcb-layout-review for the actual routing, and with design-and-review-circuit for the driving/receiving circuit topology.

Core concepts and target numbers

  • Characteristic impedance (Z0). Single-ended traces are usually targeted at 50 Ω ± 10%; differential pairs at 90–100 Ω differential (USB2 ≈ 90 Ω, HDMI/DP/PCIe/USB3 ≈ 85–100 Ω depending on spec revision — always read the exact number from the interface's own electrical spec, never assume).
  • Insertion loss (S21) and return loss (S11/S22). Specified in dB vs. frequency up to the Nyquist frequency of the link (half the baud rate for NRZ). A rough budget: total channel insertion loss should leave at least 3–4 dB of margin below what the transmitter's equalization (de-emphasis/ CTLE/DFE) can compensate; check the SerDes IP's channel-compliance spec for the actual number.
  • Eye diagram. Height (mV) and width (UI, unit interval) measured at the receiver, against the interface spec's eye mask. An eye that is "closed" (height or width below the mask) means bit errors regardless of how clean the board looks visually.
  • Jitter budget. Total jitter (TJ) at a given BER (commonly 1e-12) splits into deterministic jitter (DJ: ISI, duty-cycle distortion, crosstalk-induced) and random jitter (RJ, Gaussian, extrapolated with a Dual-Dirac model). Each stage (clock source, PCB channel, connector, cable) consumes part of the total budget — add contributions in quadrature for RJ, linearly for DJ.
  • Crosstalk. Near-end (NEXT) and far-end (FEXT) crosstalk from trace spacing, measured as coupled voltage or as an eye-closure contribution. Rule of thumb for routing: edge-to-edge spacing ≥ 3× trace width (the "3W rule") keeps coupling below roughly −40 dB for most digital traces; verify with field-solver extraction for anything speed-critical.
  • Return-path integrity. Every signal needs an unbroken reference-plane image directly underneath it. A plane split, connector transition, or via that an AC return current cannot follow (no adjacent ground via) adds an inductive discontinuity that shows up as a notch in S21/S11 and as ringing in the time domain.

Procedure

  1. Define the link budget before routing. For the interface in question, pull from the official specification (USB-IF, HDMI Forum, PCI-SIG, JEDEC for DDR, IEEE 802.3 for Ethernet):

    • Target impedance (single-ended and/or differential) and tolerance.
    • Maximum insertion loss at Nyquist, and the receiver's equalization budget.
    • The eye mask (height/width) at the measurement point (TP2/TP3 etc., per the spec's test-point definitions).
    • Maximum allowed skew (for parallel buses like DDR) or intra-pair skew (for differential pairs). Write these numbers into a hw/si/budget.md (or equivalent) before layout starts — this is the acceptance criterion the rest of this skill checks against.
  2. Compute stack-up impedance. Use a 2D field solver, not a simple parallel-plate formula, once there is soldermask, coverlay, or an adjacent plane with cutouts:

    • Open-source: scripts/stackup_solver.py in this skill wraps openEMS or the IPC-2141 closed-form equations (good to roughly ±10% for a first pass; use a full-wave solver before final release). KiCad's built-in "PCB Calculator" (Tools → PCB Calculator → Track Width tab) also gives a quick first estimate using IPC-2141/2221.
    • Commercial: Polar Si9000, Saturn PCB Toolkit (Windows, free), or the stack-up tool built into the EDA (Altium Layer Stack Manager, Cadence Allegro PCB SI). Iterate trace width/spacing and dielectric choice until the computed Z0 is within the budget's tolerance, then lock the stack-up in the fab drawing.
  3. Extract S-parameters for the critical channel. For anything running faster than ~1 Gbps per lane, or any channel with connectors/cables in the path:

    • Export the routed geometry (trace, via, connector footprint) as a 3D model and simulate in a field solver: ANSYS HFSS/SIwave, Keysight ADS (Momentum/Channel Simulator), Cadence Clarity, or the open-source openEMS / scikit-rf + a simplified via/trace model for a first pass.
    • scripts/touchstone_check.py <file.s4p> --mask references/eye-masks/<interface>.json reads a Touchstone (.sNp) file and reports insertion loss, return loss, and worst-case margin to the budget from step 1 at the link's Nyquist frequency.
    • If no field solver is available, use the IPC-2141 loss estimate from step 2 plus a fixed connector/via loss allowance (typically 0.5–1.5 dB per transition at multi-GHz rates) as a conservative substitute, and flag the result as an ESTIMATE, not a measurement.
  4. Run eye-diagram / jitter simulation. Feed the extracted S-parameters (or the estimate from step 3) into a channel simulator with the actual driver/receiver IBIS-AMI model from the transceiver's datasheet:

    • Keysight ADS Channel Simulator, Cadence SystemSI, or Mathworks SerDes Toolbox for IBIS-AMI statistical/bit-by-bit simulation.
    • scripts/eye_budget.py --insertion-loss-db <N> --rj-ps <N> --dj-ps <N> --ui-ps <N> gives a quick quadrature jitter-budget check (no IBIS-AMI) suitable for a first-pass sanity check, not a compliance-grade result. Compare the resulting eye height/width against the mask from step 1.
  5. Check crosstalk and return path on the actual layout. Export the routed board (or the critical nets) and verify:

    • No reference-plane splits crossed by a high-speed signal without a stitching capacitor or via placed within roughly 1/20 of a wavelength of the crossing.
    • Edge-to-edge spacing on coupled differential pairs and adjacent buses meets the 3W rule or better; tighten with field-solver NEXT/FEXT extraction (Si9000, HFSS) if space is constrained.
    • Every signal via that changes reference plane has an adjacent ground via within about 1 mm, so the return current has a short path. scripts/return_path_check.py <board.kicad_pcb> --nets <net-list> flags plane crossings and isolated signal vias on a KiCad board as a heuristic screen; it is not a substitute for field-solver crosstalk extraction.
  6. Validate on hardware once the board exists. A real oscilloscope or BERT measurement is the final gate:

    • TDR (time-domain reflectometry) on a VNA or a TDR-capable scope confirms the as-built impedance against the target from step 1.
    • A real-time or sampling scope with the interface's compliance test fixture gives the actual eye diagram; compare against the spec mask.
    • BERT (bit-error-rate test) at the target data rate for long enough to resolve the target BER (e.g. ≥ 10/BER unit intervals for a 1e-12 target) before declaring margin.

Data sources

  • Interface electrical specifications: USB-IF (usb.org), HDMI Forum, PCI-SIG, VESA (DisplayPort), JEDEC (DDR/LPDDR), IEEE 802.3 (Ethernet PHYs), MIPI Alliance (CSI/DSI).
  • IBIS/IBIS-AMI models from the transceiver or SerDes PHY vendor's product page (search " IBIS" or " IBIS-AMI").
  • Dielectric Dk/Df vs. frequency from the laminate datasheet (e.g. Isola, Panasonic, Rogers) — never use a single "typical" Dk across a wide frequency range for anything above a few hundred MHz; Dk/Df are frequency- dependent (dispersion).

Prerequisites

  • A routed or near-final PCB layout (KiCad, Altium, or equivalent) for the return-path and crosstalk checks; stack-up decisions in step 2 can be made pre-layout.
  • python3 with scikit-rf (pip install scikit-rf) for Touchstone analysis in scripts/touchstone_check.py. No field solver is bundled with this skill — HFSS/ADS/SIwave/Clarity or openEMS must be installed and licensed separately for S-parameter extraction in step 3.
  • The transceiver/PHY datasheet and, ideally, its IBIS-AMI model for step 4; without it, eye simulation is a rough estimate, not a compliance result.

Limitations

  • This skill estimates and screens; it does not replace a certified compliance lab for formal certification (USB-IF, HDMI, PCIe compliance all require accredited test houses with calibrated fixtures).
  • scripts/stackup_solver.py's IPC-2141 fallback is accurate to roughly ±10–15% and ignores frequency-dependent dielectric loss; treat its output as a first-pass estimate and re-verify with a full 3D field solver before committing an expensive or high-volume board to fabrication.
  • scripts/return_path_check.py is a heuristic plane/via screen on KiCad geometry; it does not compute actual crosstalk or insertion-loss numbers and can both miss real issues (e.g. a split it does not recognize) and flag benign ones (e.g. a low-speed signal crossing a split).
  • Simulation results are only as good as the model inputs: a wrong Dk/Df, a missing via stub model, or a generic (non-vendor) IBIS-AMI model can shift the predicted eye margin by several dB. Always validate on real hardware before high-volume commitment.

Quality checks before sign-off

  • Stack-up impedance within the target's tolerance, confirmed by at least one field-solver or TDR measurement (not just the IPC-2141 estimate).
  • Worst-case channel insertion loss at Nyquist, with margin to the transceiver's documented equalization budget, stated as either MEASURED (S-parameter extraction or TDR/VNA) or ESTIMATE (closed-form formula) — never left unlabeled.
  • Eye height/width at the receiver test point meets the interface's compliance mask, or the shortfall and its root cause (loss, jitter, crosstalk) is documented with a mitigation (equalization setting, stack-up change, re-route).
  • Every reference-plane crossing on a high-speed net has a stitching via or capacitor; every via on those nets has an adjacent ground via.
  • Real hardware (TDR, eye/BERT) measurement attached for any production release, not simulation alone.

示例提示词

帮我检查这条高速链路的眼图和阻抗匹配,看看能不能过兼容性测试。

所属插件

电子电气与信息工程面向电气与电子工程研究:电力系统仿真(PSS/E、PSLF、PowerWorld、pandapower、PyPSA、OpenDSS、ANDES)与越限/稳定/短路处置,FPGA 时序收敛与跨时钟域检查,电路与 PCB 设计评审,硬件产品开发流程。

安装该插件会一并获得同领域的 63 个 Skill / MCP。

同插件的其他科研 Skill

常见问题

高速信号完整性分析能做什么?

Analyzes high-speed PCB signal integrity: impedance stack-up, S-parameter extraction, eye-diagram budgeting, and crosstalk checks. Use when routing USB3/HDMI/DDR/SerDes links。

如何安装高速信号完整性分析?

在科搭 Codar 左侧点击「插件」,在「组件」页签搜索「高速信号完整性分析」并安装;在 Claude Code 或 Codex 中可以发送本页的安装提示词,或下载 Skill 包后解压到 ~/.claude/skills 或 ~/.codex/skills。

高速信号完整性分析需要哪些权限?

该科研 Skill没有声明额外权限;执行具体操作时仍由 Agent 按自身规则请求确认。

高速信号完整性分析经过官方审核吗?

是。SpeedSkills 收录的每个版本都会记录内容摘要与权限声明,当前版本已通过安全审核。