LVDS/HCSL/LVPECL Fanout Buffers for PCIe/SerDes Clocks
← Back to:Reference Oscillators & Timing
LVDS/HCSL/LVPECL fanout buffers replicate one clean clock into many low-skew outputs—so every endpoint sees the same electrical contract and stays inside the system jitter budget. The practical outcome is predictable termination, routing, and verification: control where the clock is measured, how it is terminated, and how additive jitter + board noise are proven against pass criteria.
H2-1. Definition & Page Boundary: What a Fanout Buffer Does (and Doesn’t)
A fanout buffer is a low-additive-jitter, low-skew distribution stage that replicates one clock into N matched outputs using defined electrical standards (LVDS/HCSL/LVPECL). It improves clock-tree reliability by isolating loads and enforcing consistent edge quality at multiple endpoints.
Covers (this page):
- Fanout role in a clock tree: replicate + match (1→N or 2→N), not frequency synthesis.
- What matters in practice: additive jitter, channel-to-channel skew, output standards, termination, routing, power integrity, and measurement setup.
- Engineering hooks for PCIe/SerDes refclock distribution (electrical contract only).
Does NOT cover (avoids topic overlap):
- Jitter cleaning / loop behavior (PLL cleaners, attenuators): belongs to the “Jitter Attenuators / Clock Cleaners” page.
- Hitless switching, clock muxing, redundancy state machines: belongs to “Glitch-Free Clock Mux”.
- Dynamic routing matrices and guard/bypass paths: belongs to “Clock Crosspoint Switch”.
- Phase alignment and ps–ns trims: belongs to “Programmable Delay / Phase”.
- Protocol/system tutorials for PCIe/JESD204/SerDes: belongs to their interface-focused pages (only electrical constraints are referenced here).
When to use a fanout buffer:
- One clock must drive multiple endpoints where direct parallel wiring would create stubs, reflections, or uncontrolled loading.
- Endpoints require consistent output levels (LVDS/HCSL/LVPECL) and a repeatable termination strategy.
- System needs controlled skew between branches (multi-domain FPGA/SoC, multi-SerDes, multi-converter timing trees).
- Clock-tree verification requires a clean point to measure buffer contribution separately from board-induced degradation.
Minimum dictionary (used consistently on this page)
- Additive jitter: jitter added by the buffer itself (use the vendor’s stated integration method/window).
- Integrated RMS jitter: RMS time jitter over a defined band/window; valid only when window is stated.
- Channel-to-channel skew: output-to-output time offset (static + drift across voltage/temperature).
- Output standard: LVDS/HCSL/LVPECL electrical levels + common-mode expectations + termination style.
- Termination: impedance/return strategy that controls reflections and common-mode behavior at the receiver.
- Endpoint: a consumer of the clock (PCIe/SerDes/FPGA/converter) with its own jitter tolerance contract.
H2-2. Where Fanout Buffers Fit in Real Clock Trees
The practical clock-tree choice is rarely “which part number,” but “which distribution topology.” A single-stage fanout (one device drives all endpoints) is simplest, while a two-stage fanout (one device feeds local fanouts) often wins when endpoints are physically spread out or cross connectors and zones.
Single-stage fanout (1→N):
- Best when: endpoints are in one zone/cluster, routes are short, and the board has a clean return path.
- Engineering benefit: one distribution point, consistent output standard, fewer supplies and fewer layout hotspots.
- Main risks: long routes and stubs increase reflections and EMI; skew variation grows with length/return discontinuities.
- Verification hook: compare output quality at the fanout pins vs at the far endpoints to separate buffer vs board effects.
Two-stage fanout (1→K→N):
- Best when: endpoints span multiple zones, cross connectors/cards, or require separate routing/ground partitions.
- Engineering benefit: shorter local routes reduce reflection sensitivity; distribution is “localized” and easier to shield.
- Main risks: extra devices add their own additive jitter; power integrity and isolation must be enforced at each stage.
- Verification hook: stage-by-stage acceptance (Stage-1 outputs pass before Stage-2 validation begins).
Placement strategy (where to put the fanout):
- Near the source/cleaner: consistent electrical environment; may require longer endpoint routes.
- Near the endpoints (zonal fanouts): short endpoint routes; requires clean “trunk” routing to each zone.
- Rule of thumb: prioritize short, well-controlled routes after the final fanout stage; avoid branching stubs at endpoints.
Redundancy concept (main/backup clocks):
- Switch before fanout: a single switch point controls all endpoints; faults propagate widely if switching is imperfect.
- Switch after fanout (per-zone/per-endpoint): local containment; higher routing and validation complexity.
- Boundary note: hitless switching mechanics belong to the “Glitch-Free Clock Mux” page; this page focuses on fanout placement trade-offs.
Decision checklist (choose topology with measurable criteria)
- If endpoints are spread across multiple physical zones (or cross connectors), prefer two-stage (zonal fanouts).
- If the route from fanout to endpoints would create long stubs or uncontrolled branching, move to zonal fanout.
- If skew must remain stable across temperature/voltage, minimize route length differences after the final fanout stage.
- If EMI margin is tight, shorten high-edge-rate routes and avoid return-path discontinuities; two-stage often reduces the hardest routes.
- If measurement shows endpoint jitter is dominated by board effects, improve routing/termination before chasing lower buffer specs.
- If supplies are noisy near endpoints, keep the final fanout close to a clean power island or enforce stricter decoupling locally.
- If output standards differ (LVDS vs HCSL vs LVPECL), group endpoints by standard and avoid mixed termination assumptions.
- If validation time is limited, choose a topology that allows stage-by-stage pass/fail isolation.
H2-3. Output Standards Deep Dive: LVDS vs HCSL vs LVPECL (Engineering View)
LVDS, HCSL, and LVPECL are not “interchangeable differential clocks.” Each defines a different electrical contract: driver behavior, swing/common-mode expectations, and how termination must be implemented to avoid reflections and common-mode issues.
LVDS (low swing, 100Ω differential)
What it wants
- Clean 100Ω differential environment and continuous return path.
- Receiver input common-mode compatibility (avoid “hidden” CM range violations).
- Minimal stubs; route as a true differential pair (controlled impedance).
Typical termination
- 100Ω differential at the receiver (or at the far end of the line).
- AC coupling is used only when DC common-mode/bias contracts cannot be met directly.
Common pitfalls
- Parallel “tee” wiring (one output feeding multiple receivers) creates stubs and reflections.
- Common-mode mismatch between driver and receiver reduces margin and can degrade eye quality.
- Probing at the wrong point (or loading the line) looks like “jitter,” but is measurement artifact.
HCSL (PCIe-style, termination placement sensitive)
What it wants
- Receiver-side termination consistent with the platform contract (often near the receiver).
- Controlled differential routing; avoid long branches after the last split.
- Stable reference/return path; common-mode behavior is part of the compliance story.
Typical termination
- Receiver-side termination network (platform-defined; avoid “creative” mid-line terminations).
- Keep the termination physically close to the receiver pins to prevent stub-induced ringing.
Common pitfalls
- Termination placed at the wrong end creates overshoot/ringing that looks like duty/jitter problems.
- Assuming “any 100Ω” is enough without following the receiver’s contract.
- Using one HCSL output to feed multiple endpoints without a proper clock-tree stage.
LVPECL (higher swing, bias/common-mode must be respected)
What it wants
- Defined bias/common-mode path (receiver must see the expected operating region).
- Termination that both controls reflections and sets the correct DC operating point.
- Short, clean differential routing to reduce sensitivity to return-path discontinuities.
Typical termination
- Split termination to Vtt (or an equivalent bias network) near the receiver.
- AC coupling is used when DC bias must be re-established at the receiver side.
Common pitfalls
- Missing bias path (or wrong Vtt) shifts the operating point and causes waveform distortion.
- “Looks fine no-load” but collapses when connected (termination/bias assumptions violated).
- Mixed-standards bring-up without a compatibility check (DC and CM requirements conflict).
Compatibility quick check (before choosing a termination)
- If driver and receiver common-mode/bias contracts do not match, consider AC coupling and re-bias at the receiver.
- If a clock crosses a connector or zone boundary, lock down termination location and avoid mid-line stubs.
- If overshoot/ringing changes strongly with termination placement, suspect stub length and return-path discontinuity.
- If only some outputs fail, verify per-output loading, adjacency coupling, and local supply/return integrity.
- Do not compare “jitter” across parts without matching termination, routing, and measurement setup.
Note: Interface/protocol details (PCIe/JESD204/SerDes) are handled on their dedicated pages; this page stays at the electrical/termination layer.
H2-4. Key Specs That Actually Matter: Additive Jitter, Skew, and Output Quality
Datasheet numbers only help when they map to a system budget. For fanout buffers, the “make-or-break” specs are additive jitter (with a stated measurement window), channel-to-channel skew (including drift), and output quality that remains compliant under the intended standard and termination.
Additive jitter (RMS / phase jitter)
Metric
Jitter added by the fanout stage itself. The value is meaningful only when the vendor states the integration window/band and conditions.
Why it matters
It consumes endpoint jitter tolerance. In a clock tree, it must be counted once (avoid double counting with upstream/downstream measurements).
How to measure
- Compare a direct input reference vs fanout output under identical termination and probe strategy.
- Use a consistent jitter window (or PN integration band) aligned to endpoint requirements.
Guardband
- Require the window/band to be explicitly stated; ignore “jitter” numbers without it.
- Use worst-case/max specs (or add margin) across voltage/temperature and intended output standard.
Channel-to-channel skew (static + drift)
Metric
Output-to-output time offset. Separate static skew (at one condition) from drift across PVT (power/voltage/temperature) and aging.
Why it matters
Static skew affects alignment at bring-up; drift consumes timing margin over temperature and can create “intermittent” failures in multi-domain systems.
How to measure
- Measure multiple outputs simultaneously with matched probing and consistent thresholds.
- Repeat across temperature points to capture drift (not only room-temperature static).
Guardband
- Budget for both buffer skew and routing-induced skew; keep route length differences minimal after the final fanout stage.
- Use drift-aware limits (static-only “typ” is insufficient for robust systems).
Output quality (rise/fall, duty-cycle distortion, compliance)
Metric
Edge rate and duty-cycle behavior under the intended output standard and termination. Compliance depends on load and routing.
Why it matters
- Too fast: EMI and reflection sensitivity rise; return-path discontinuities become dominant.
- Too slow: crossing uncertainty increases and effective timing margin shrinks.
- DCD can disturb downstream recovery circuits even when “frequency is correct.”
How to measure
- Measure at the receiver location with the real termination in place.
- Verify swing/common-mode stays in contract over PVT and across all outputs.
Guardband
- Evaluate worst-case load and routing (connectors, long traces, via transitions).
- Prefer parts with clear compliance statements under stated conditions.
Output isolation / crosstalk (multi-output reality)
Metric
How much one output’s activity perturbs another output via internal coupling, shared supplies, or adjacent routing.
Why it matters
It creates “some channels fail, others pass” behavior. It also makes endpoint jitter appear workload-dependent.
How to measure
- Toggle or heavily load one output while monitoring another output’s phase/edge stability.
- Repeat with different adjacency groupings (neighbor pins vs far pins).
Guardband
- Group sensitive endpoints on outputs with better isolation (if the package/pinout provides it).
- Enforce local decoupling and routing separation to reduce coupling paths.
Supply sensitivity (PSRR is not a shield)
Metric
How supply/ground noise turns into edge timing noise and/or compliance failures at the outputs.
Why it matters
A “good jitter part” can perform poorly on a noisy island. Supply-induced time noise often dominates after routing is fixed.
How to measure
- Compare output jitter with clean vs stressed supply conditions (controlled ripple injection if available).
- Correlate jitter changes with measured supply noise near the device pins.
Guardband
- Budget for worst-case supply noise and enforce local decoupling/isolation requirements.
- Validate at the board level; datasheet “typ” is not a system guarantee.
H2-5. Signal Integrity & Termination Design: Making the Waveform Real
Board-level “jitter” is often waveform integrity: reflections, stubs, common-mode/bias violations, or measurement loading. The goal is to turn scope observations into a repeatable diagnosis loop: identify the symptom class, validate the dominant root cause with a quick check, then apply a minimal, verifiable fix.
Symptom triage (classify before changing parts)
- Large overshoot/undershoot: impedance step, termination mismatch, return-path discontinuity.
- Ringing with slow decay: stub/branch resonance, connector discontinuity, termination too far from the receiver.
- Multiple threshold crossings / “double edge”: reflection + crossing uncertainty; often worsened by probing or a mid-line tee.
- “Jitter looks huge” but frequency is correct: crossing noise from ringing/common-mode movement; measurement artifacts are common.
- One output looks clean, another is unstable: multi-output coupling paths (handled in H2-6) or per-output loading differences.
Root-cause map (most common, highest leverage)
Impedance discontinuity
Via fields, layer swaps, connectors, and reference-plane changes create Z steps that reflect energy and distort edges.
Stub / branch / test point
Any tee after the last fanout creates a reflection source. Even “small” stubs can cause ringing and multiple crossings.
Termination placed too far away
A termination that is not at the receiver behaves like a stub. It can reduce, not eliminate, ringing at the receiver.
Common-mode / bias contract broken
Especially in HCSL/LVPECL-style environments: the receiver needs the expected bias/common-mode path to stay in spec.
Return-path discontinuity
Crossing splits/slots forces return current detours, raising loop inductance and turning edge energy into ringing.
Probe and measurement loading
Probe capacitance/ground leads change the circuit. A “bad waveform” can be created by the measurement setup.
Termination placement decision (receiver-end by default)
- Default: place termination at the receiver (or the far end). This is the most robust way to absorb reflections where they matter.
- Branch-point termination (limited cases): only when the post-branch segments are very short and symmetric, and the branch point must define the impedance/bias environment.
- Hard rule: do not use one fanout output to feed multiple receivers (“one-to-many tee”). That is a clock-tree topology problem (H2-2), not a termination trick.
AC coupling & DC bias (use for CM/bias compatibility, not as a ringing cure)
- Use AC coupling when: driver and receiver common-mode/bias contracts cannot be satisfied directly, or when a receiver must establish its own bias.
- Bias must exist somewhere: AC caps block DC; the receiver side must provide the required bias/common-mode path (e.g., Vtt network).
- Do not expect: AC coupling to “solve reflections.” Reflection control still depends on impedance continuity, stub control, and termination placement.
Fast checks (quick experiments that confirm the dominant cause)
- Move the measurement point: compare waveform at the receiver vs near the source to separate routing artifacts from device behavior.
- Change the probe strategy: reduce probe loading, shorten ground return, or use a differential probe where appropriate.
- Temporarily enforce receiver-end termination: add or relocate the 100Ω (or the contract-defined network) to the receiver end.
- Eliminate the stub: remove a test pad/fly-wire branch or cut a tee to see if ringing/multiple crossings vanish.
- Toggle AC coupling (diagnostic): only to validate a CM/bias hypothesis; confirm that receiver bias remains valid.
- Return-path check: if the route crosses a split/slot, add a controlled return stitch (for diagnosis) and re-measure.
Fix actions & pass criteria (keep it testable)
Stub-driven ringing Fix
- Remove/shorten the tee and keep a single point-to-point link after the last fanout.
- Move termination to the receiver end and keep the termination physically close to the pins.
Pass: ringing no longer creates multiple threshold crossings at the receiver.
Impedance steps / connector discontinuity Fix
- Reduce discontinuities (fewer layer swaps, controlled connector breakouts, consistent reference planes).
- Add stitching vias across transitions to maintain a continuous return path.
Pass: overshoot/undershoot reduces and edge shape becomes monotonic at the receiver.
Common-mode / bias issues Fix
- Ensure the receiver sees the required bias/common-mode path (Vtt or contract-defined network).
- Use AC coupling only when the receiver re-establishes a valid DC operating point.
Pass: swing/common-mode stays within contract under the real termination and load.
H2-6. Crosstalk, Channel Isolation, and Multi-Output Reality
In multi-output fanout systems, “one output passes, another fails” is a signature of coupling paths: internal channel coupling, shared supply/return impedance, and routing electromagnetic coupling. Isolation is a system property: pin grouping, power/return strategy, and physical routing all contribute.
Coupling paths (where the “difference between outputs” comes from)
IC-level coupling
- Neighbor outputs can couple through internal structures and shared substrate paths.
- Heavy loading or aggressive switching on one output can modulate edge stability on another.
Power coupling
- Shared VDD impedance converts output switching current into local supply movement.
- Supply-induced time noise can dominate after routing and termination are “correct.”
Return-path and routing coupling
- Parallel routing, layer adjacency, and missing via fences raise field coupling.
- Crossing plane splits forces shared/detoured returns, coupling channels through the return network.
Isolation strategy: neighbor outputs vs far outputs (grouping and guarding)
- Group by sensitivity: assign the most sensitive endpoints to outputs with the best isolation (avoid placing them on adjacent pins when possible).
- Separate noisy loads: isolate heavily loaded outputs and keep their routes away from sensitive pairs.
- Guard the routes: increase spacing, add ground guarding, and use via fences near connectors and long parallel segments.
- Partition returns: keep the return path for noisy outputs from crossing through sensitive clock areas.
- Verify with injection: stress one output (toggle/heavy load) and observe whether another output’s edge stability changes.
Do / Don’t checklist (engineering actions)
Do
- Keep sensitive clock pairs short and away from high-switching digital aggressors.
- Use consistent reference planes and stitch returns across transitions.
- Place decoupling close to the fanout supply pins and keep loop area small.
- Use spacing/guarding/via fences in long parallel routing regions.
- Group outputs so that adjacent channels are not both “critical.”
Don’t
- Do not route multiple outputs in long, tight parallel bundles without guarding.
- Do not cross plane splits/slots with differential clock pairs without a planned return strategy.
- Do not share noisy supply islands with sensitive outputs without isolation/decoupling discipline.
- Do not assume “all outputs are identical” under different loads and adjacency conditions.
- Do not diagnose per-output differences using different measurement setups (probe/ground/threshold mismatch).
H2-7. Power Integrity & Noise Coupling: PSRR Is Not a Magic Shield
A fanout buffer can meet low additive jitter on paper and still show unstable edges on a real PCB. The common reason is power/return noise converting into edge timing noise: supply ripple or ground bounce shifts the effective threshold/crossing point, so the time of the edge moves even when frequency is unchanged.
The engineering chain: supply/return noise → crossing shift → time jitter
- Supply ripple modulates swing / driver threshold behavior and reduces crossing determinism.
- Ground bounce moves the reference seen by the receiver, turning return impedance into time noise.
- Edge slope matters: the slower the edge at the receiver, the more a small voltage disturbance becomes a larger Δt.
Practical takeaway: improving the device alone may not help until the local supply/return impedance is reduced and controlled.
Coupling channels that dominate real boards
A) Supply ripple → threshold/swing modulation
- Symptom: edge stability changes with rail noise; output-to-output behavior varies with load.
- Quick check: compare two conditions (cleaner local supply vs intentionally worse local impedance) and look for strong correlation.
- Fix direction: strengthen the local decoupling loop and isolate noisy upstream rails.
B) Ground bounce → reference movement
- Symptom: edges degrade when nearby outputs or aggressors switch; jitter appears “bursty.”
- Quick check: improve return continuity (stitching/shorter loops) and observe whether instability reduces.
- Fix direction: continuous reference planes + return stitching near transitions and connectors.
C) Filtering choices are not free (FB/π)
- Symptom: adding a bead or π filter changes behavior by frequency; improvements in one condition can regress in another.
- Quick check: compare “strong local decoupling only” vs “bead/π + weak local” to separate root cause.
- Fix direction: only add isolation once the local loop is already tight and stable.
D) Multi-rail devices: AVDD/DVDD/OVDD allocation
- Symptom: one rail dominates sensitivity; “datasheet jitter” cannot be reproduced.
- Quick check: harden rails one-by-one and observe which improvement has the largest edge-stability gain.
- Fix direction: keep rail intent clear (output driver vs digital core) and isolate noise sources accordingly.
Minimal power network template (recommended decoupling stack)
Template Hierarchy
- Upstream rail → (optional: LDO/RC) → (optional: bead/π) → local cap stack → device pins → tight return path
- For multi-rail parts: treat each rail as a separate loop (local caps + return) before adding shared isolation.
Principles Do
- Place high-frequency decoupling closest to the pins, minimizing loop area and via count.
- Use a “stack”: HF (closest) → MF (nearby) → bulk (support) so the local impedance stays low over a wide band.
- Keep the return continuous; a great capacitor with a bad return path still produces time noise.
Failure modes Avoid
- Bead/π used as a substitute for poor local decoupling (local loop remains high impedance).
- Too much isolation without a stable local network (frequency response moves noise into sensitive regions).
- Return detours (splits/slots/long ground paths) that turn switching current into edge movement.
Pass criteria (keep it verifiable)
- Correlation breaks: edge stability no longer tracks upstream rail activity or nearby switching events.
- Repeatability improves: the same output shows similar edge behavior across power states and board temperature gradients.
- Local dominance: strengthening local decoupling yields diminishing returns (indicating the local loop is no longer the bottleneck).
H2-8. PCB Layout & Routing Cookbook (Differential Clocks)
Differential clocks are not “immune by default.” Consistent impedance, continuous reference planes, short routes, minimal vias, and receiver-end termination determine whether the clock delivered to endpoints matches the intended edge quality and skew budget.
Top 12 rules (each rule = one action + one reason)
- 1) Control impedance first: define differential impedance from the stackup and keep it consistent to suppress reflections.
- 2) Keep geometry constant: avoid sudden width/spacing changes that create local Z steps.
- 3) Match lengths to X: match intra-pair to Xdiff and inter-channel to Xch derived from the system skew budget.
- 4) Route short and straight: reduce path uncertainty and loss; clocks do not benefit from “scenic routes.”
- 5) Minimize vias/layer swaps: each transition adds discontinuity and return-path stress.
- 6) Avoid stubs after the last split: tees and test stubs are reflection sources and create multiple crossings.
- 7) Keep the reference plane continuous: a broken return path converts switching current into edge movement.
- 8) Do not cross splits/slots: return detours raise loop inductance and produce ringing and time noise.
- 9) Stitch returns at transitions: add stitching vias near layer changes and connectors to maintain return continuity.
- 10) Terminate at the receiver: place the termination network physically close so it does not become a stub.
- 11) Create a clock keepout zone: keep clocks away from switching power nodes, hot loops, and strong magnetic fields.
- 12) Guard long parallel regions: increase spacing and add guarding/via fences to reduce coupling.
Placement notes (what matters most at the pins)
- Receiver-end termination: place the network as close as possible to receiver pins to ensure the receiver sees the intended impedance.
- Fanout neighborhood: keep the supply/return loops tight, and avoid placing high-current switching nodes adjacent to the clock outputs.
- Connectors: treat as discontinuities; keep breakout symmetric and maintain reference continuity with stitching vias.
Quick self-check before fab (8 items)
- Any clock pair crosses a plane split/slot (yes/no)?
- Any tee/stub exists after the last fanout (yes/no)?
- Termination is physically at the receiver (yes/no)?
- More than a few vias on a critical clock path (yes/no)?
- Long parallel clock runs without spacing/guarding (yes/no)?
- Return stitching near transitions/connectors is missing (yes/no)?
- Clock routes pass near switching power hot loops (yes/no)?
- Critical and noisy outputs are grouped on adjacent pins/routes (yes/no)?
H2-9. Control & Behavioral Features: OE, Output Config, SSC Passing, Startup
Control pins define real-world behavior: enable/disable transitions, default states, and configuration latching can create phase discontinuity or edge artifacts even when steady-state jitter specs look excellent. Treat OE and mode selection as part of the electrical contract and the system startup timeline.
Output Enable (OE): not just ON/OFF
Typical OE behaviors (engineering view)
- Asynchronous tri-state / force-state: output changes immediately with OE; highest risk of runt pulses if OE toggles near edges.
- Synchronous gating: OE is applied on an internal boundary; lower glitch risk but can introduce a deterministic phase step.
- Glitch-reduced (not guaranteed glitch-free): common in practice; still needs a safe enable window.
Engineering risks to manage
- Phase continuity: the first valid edge after OE may not align with the previous phase relationship.
- Extra crossings: enable/disable can create short pulses or double edges that trigger downstream re-lock or false clocks.
- Disabled-state contract: Hi-Z vs forced state can affect termination/bias networks and receiver behavior.
Output configuration: switching the electrical contract
- Multi-standard / multi-swing outputs: changing mode changes swing, common-mode expectations, and termination compatibility.
- Strap/latch timing matters: if mode pins are sampled at power-up, unstable straps can cause wrong output behavior until reset.
- Group/bank implications: mixed endpoint sensitivity benefits from grouping (critical outputs not adjacent to noisy/heavy loads).
SSC passing: “transparent” depends on whether the clock is regenerated
Decision rule (practical)
- Pure buffer/re-driver: more likely to pass SSC shape through to outputs.
- Regenerating/PLL-based behavior: may attenuate or reshape SSC depending on internal dynamics.
What matters to endpoints
- Downstream tolerance: some CDR/PLLs are sensitive to SSC profile changes.
- Verification focus: confirm “lock/training stability under worst-case SSC,” not just “SSC exists.”
Scenario cards (behavior-focused)
Cold start
Likely risk: unstable straps/rails → wrong mode or early OE glitch.
Quick check: observe first edges after reset/OE at endpoint.
Fix: hold OE inactive until rails + configuration are stable.
Pass: no extra crossings; endpoint reaches stable lock/training once.
Warm restart
Likely risk: phase discontinuity when clocks resume.
Quick check: compare pre/post restart edge behavior at endpoint.
Fix: restart inside a system-defined “ignore window” or re-train window.
Pass: deterministic restart behavior; no unexpected re-lock loops.
Hot-plug
Likely risk: sudden loading/termination change → edge artifacts.
Quick check: monitor unaffected ports during plug events.
Fix: isolate ports, stage enable, and keep stubs/tees out of live paths.
Pass: other endpoints remain stable; plugged endpoint trains normally.
Fault recovery
Likely risk: brown-out causes partial reset; outputs may glitch.
Quick check: correlate rail events with clock instability.
Fix: enforce a full, deterministic sequence: reset → config → OE.
Pass: one clean recovery path; no repeated re-lock oscillations.
H2-10. PCIe / SerDes Reference Clock Distribution Hooks (Electrical Contract Only)
PCIe/SerDes reference clock distribution is a clock-tree and electrical contract problem: output standard (often HCSL), receiver-end termination, common-mode expectations, and a jitter/skew budget that must hold at the endpoint. This section avoids protocol details and focuses on engineering constraints a fanout buffer must satisfy.
SRNS vs SRIS: engineering impact on the clock tree (no protocol)
Shared-reference style (SRNS-like)
- Tree implication: one clean source → fanout → multiple endpoints; channel-to-channel skew consistency becomes critical.
- What to watch: consistent termination/common-mode per endpoint and stable behavior under SSC if present.
Independent-reference tolerant style (SRIS-like)
- Tree implication: endpoints may tolerate more independence; per-link absolute edge quality and endpoint sensitivity dominate planning.
- What to watch: avoid assumptions—verify endpoint stability under the intended clock profile and power/thermal corners.
Practical interpretation: regardless of SRNS/SRIS framing, the endpoint electrical contract (jitter/skew/termination/common-mode) must be met at the receiver pins.
HCSL hooks (contract essentials)
- Termination at the receiver: treat placement as part of the contract; “mid-line termination” behaves like a stub.
- Common-mode consistency: ensure the receiver sees the intended bias/common-mode environment (especially across connectors and card edges).
- AC coupling only when needed: use it to satisfy bias/common-mode constraints, not as a substitute for routing quality.
Endpoint classes (sensitivity-driven grouping)
RC / CPU
Cares about: stable reference quality and consistent distribution.
Tree hook: place on the cleanest, shortest, most repeatable branch.
Verify: stable crossings and no training/re-lock loops across power states.
Switch
Cares about: multi-port consistency; skew and distribution symmetry.
Tree hook: group ports and manage skew within Xch.
Verify: consistent behavior across ports under the same clock profile.
Retimer
Cares about: input quality; can amplify marginal refclock behavior.
Tree hook: isolate from noisy branches and avoid routing discontinuities.
Verify: lock stability under worst-case SSC/power/thermal corners.
Endpoint
Cares about: receiving a valid electrical contract at its pins.
Tree hook: avoid tees/stubs; keep termination at the receiver.
Verify: stable crossings; no sensitivity to neighboring switching events.
Design checklist for PCIe refclk fanout (≤10)
- Use the intended output standard and keep the termination contract at the receiver.
- Maintain common-mode/bias expectations end-to-end (including connectors and card edges).
- Keep the route point-to-point after the last fanout output (no tees/stubs).
- Minimize vias and keep reference planes continuous along the route.
- Plan channel-to-channel skew to stay within Xch (system-derived).
- Plan total jitter budget: source + fanout additive + board contribution ≤ X.
- Group critical endpoints away from noisy/heavy-load branches.
- Validate behavior under worst-case SSC usage (if SSC is present in the system).
- Validate across power/thermal corners; do not rely on typical-only conditions.
- Keep OE/control sequencing deterministic for restart and recovery.
H2-11. Verification & Measurement: How to Prove Jitter/Skew Without Lying to Yourself
Jitter and skew numbers are only meaningful when definitions, bandwidth, reference choices, and measurement points are controlled. This section standardizes what to measure, how to measure it, which traps create “fake-good” or “fake-bad” results, and how to write pass criteria that trace back to a system budget.
A) What to measure (definitions + mandatory metadata)
RMS jitter (time-domain)
Use: budgeting and comparing devices when bandwidth/filters are identical.
Common misuse: comparing RMS jitter values measured with different integration bands or different points on the tree.
Must report: method (TIE/period), bandwidth/integration range, thresholding, measurement point, termination condition.
TIE (Time Interval Error)
Use: correlating endpoint behavior to a reference and spotting deterministic patterns.
Common misuse: reference clock is polluted (instrument or system), then TIE blames the DUT.
Must report: reference source, recovery method, trigger strategy, timebase stability.
Period / cycle-to-cycle jitter
Use: quick sensitivity checks to edge-quality and measurement noise.
Common misuse: insufficient bandwidth or heavy filtering makes results look artificially small.
Must report: scope bandwidth, probe type/loading, filtering/averaging settings.
Phase noise → integrated jitter
Use: defensible, comparable reporting when integration bounds are standardized.
Common misuse: changing offset integration limits changes jitter without changing the DUT.
Must report: PN offset range, integration limits, instrument floor / correlation setting (if used).
Skew (channel-to-channel)
Use: verifying multi-endpoint alignment and tolerance to drift.
Common misuse: different cable/fixture delays masquerade as DUT skew.
Must report: identical measurement points, identical terminations, fixture delay calibration, temperature/power state.
Rule: Any jitter/skew number without the metadata above is a non-comparable number (not safe for selection or sign-off).
B) How to measure (two primary routes)
Route 1: Oscilloscope time analysis (TIE / period)
- Best for: bring-up comparisons, fast A/B checks, trap isolation.
- Key controls: bandwidth, probe loading, trigger strategy, reference quality, sample depth.
- Blind spots: instrument timebase/trigger noise can dominate low-jitter sources.
Route 2: Phase-noise analyzer / SSA (PN curve + integration)
- Best for: budget closure, defensible reporting, root-cause by offset region.
- Key controls: offset range, integration bounds, RBW/VBW, averaging, instrument noise floor.
- Blind spots: wrong coupling/termination or wrong measurement point still produces misleading “nice” curves.
C) Measurement traps (symptom → isolation → fix)
Trap 1: probing a stub (wrong point)
Trap 2: probe loading and return path
Trap 3: insufficient bandwidth / hidden filtering
Trap 4: reference contamination
Trap 5: skew measured with uncalibrated fixture delays
D) Pass criteria: derive X from the system budget (not guesswork)
Use placeholder thresholds (X, Xch) in the text, but compute them from a documented budget so results remain portable across boards and revisions.
Reporting rule: state the measurement method and integration bounds that X applies to (TIE/period vs PN-integrated jitter).
E) Test templates (Quick / Deep / Production)
Quick check (bring-up)
- Goal: fast A/B comparisons after layout or SI changes.
- Method: scope-based edge + TIE/period checks at the endpoint contract point.
- Output: qualitative pass/fail + trend (better/worse) with metadata recorded.
Deep check (lab proof)
- Goal: budget closure and defensible reporting.
- Method: PN curve + integrated jitter using standardized offset bounds; cross-check with time-domain at the same point.
- Output: numbers tied to the budget back-solve method and worst-case conditions.
Production check (factory)
- Goal: fast, repeatable screening within test-time limits.
- Method: reduced set of checks with a consistent fixture; verify that the fixture delay is stable.
- Output: pass/fail using X derived from the system budget with an added guardband.
H2-12. Engineering Checklist + Applications + IC Selection Notes
This section compresses the page into an execution pack: a pre-design checklist, schematic/PCB checks, bring-up and production hooks, plus a selection flow that forces budget alignment and a verification plan (instead of selecting by a single “typical jitter” number).
H2-12A. Engineering checklist (primary)
Before design
- Standard + endpoints: LVDS / HCSL / LVPECL output contract must match receiver expectations.
- Frequency plan: nominal and corner frequencies; include SSC presence/absence as a system attribute.
- Output count (N) + tree topology: avoid “one output driving multiple loads” after the last fanout stage.
- Budget definition: pick one jitter/skew reporting method and integration bounds for all comparisons.
Schematic contract
- Termination & bias: place termination at the receiver; ensure common-mode/bias contract is met.
- AC coupling: use only when it is needed to satisfy bias/common-mode constraints.
- OE default state: power-up must be deterministic; avoid uncontrolled edges before configuration is stable.
- Straps & selects: define pull networks so mode pins are stable during latch windows.
PCB implementation
- Differential routing: continuous reference plane, minimal vias, no stubs/tees after fanout outputs.
- Isolation: keep clock region away from switching power inductors and high di/dt loops.
- Decoupling: local high-frequency MLCC at each supply pin group, tight loop area.
Bring-up
- Measurement points: design a safe, non-stub measurement access path (launch/tap) near the contract point.
- Instrument metadata: record bandwidth, trigger strategy, reference source, and integration bounds.
- Acceptance: derive X and Xch from the system budget with guardband (see H2-11).
Production
- Test time: pick a reduced check set that is stable and fixture-friendly.
- Fixture consistency: verify cable/fixture delay stability; re-check periodically.
- Corner sampling: plan temperature/voltage sampling for worst-case drift and skew.
Note: Any example part number below is a starting point only. Always verify package, suffix, value, tolerance, and availability against the current BOM rules.
H2-12B. Applications (boundary-safe, typical scenarios)
PCIe / SerDes multi-endpoint refclk distribution
Why fanout: consistent electrical contract across many endpoints with controlled skew and additive jitter.
Multi-FPGA / multi-card clock distribution
Why fanout: structured clock-tree stages that reduce loading, isolate reflections, and keep routing short near endpoints.
Test/ATE/bench platforms with many reference outputs
Why fanout: repeatable output levels and predictable enable/restart behavior across multiple instruments and fixtures.
H2-12C. IC selection notes (logic + concrete part numbers)
Selection inputs (define the contract first)
- Output standard: LVDS / HCSL / LVPECL (must match endpoint contract).
- N outputs + frequency range: include worst-case operating corners.
- Additive jitter target: expressed in the same measurement method and integration bounds used in verification.
- Skew target: channel-to-channel Xch derived from system alignment tolerance.
- Power rails + package: rail partitioning, decoupling feasibility, thermal constraints.
- Behavioral needs: OE behavior, default state, mux/select pins, SSC passing expectations (if SSC exists in the tree).
Reference fanout IC examples (starting points only)
PCIe-style HCSL clock buffers (fanout)
- Diodes PI6CBF18501 — 5-output PCIe clock buffer (HCSL), with on-chip terminations (verify gen/feature set and conditions).
- Diodes PI6CB332001 — 20-output PCIe clock buffer (HCSL), on-chip terminations (verify output count fit and power/thermal constraints).
LVDS clock fanout buffers
- TI CDCLVD1204 — 2:4 low additive jitter LVDS buffer (input mux, LVDS outputs).
- TI CDCLVD1208 — 2:8 low additive jitter LVDS buffer (input mux, LVDS outputs).
- TI CDCLVD2102 — dual-input distribution to multiple LVDS output pairs.
- TI SN65LVDS104 — 1:4 LVDS clock fanout buffer.
LVPECL fanout buffers / translators
- TI LMK00725 — low-skew fanout buffer providing multiple 3.3V LVPECL outputs (verify input type and enable behavior).
- onsemi NB3N853531E — 1:4 LVPECL fanout buffer with selectable input (includes crystal/CMOS input path; verify usage mode).
- onsemi NB3N853501E — 2:1 mux to 1:4 LVPECL fanout buffer outputs (verify mux/control expectations).
- Diodes PI6C485311 — low-skew LVPECL fanout buffer with selectable differential inputs (verify translation/termination constraints).
Selection warning: Avoid comparing “typical additive jitter” across parts unless the integration bounds, test setup, and measurement points match (standardize via H2-11).
Support BOM examples (termination / coupling / PI filtering)
These example material numbers speed up sourcing; verify value, size, temp rating, and procurement rules.
- 100Ω differential termination resistor: Yageo RC0402FR-07100RL (0402, 1%, example).
- 49.9Ω single-ended term / bias resistor: Yageo RC0402FR-0749R9L (0402, 1%, example).
- 0.1µF decoupling MLCC (per pin group): Murata GRM155R71C104KA88D (0402, X7R, example).
- 1µF local bulk MLCC (per rail region): Murata GRM188R71C105KA12D (0603, X7R, example).
- Ferrite bead for supply isolation: Murata BLM18AG601SN1D (0603, 600Ω@100MHz, example).
- Low-noise LDO (clock rail option): TI TPS7A47 or ADI LT3042 (use only if the rail noise budget requires it; verify stability and thermal).
Datasheet fields that often mislead
- Typical vs max: sign-off should use worst-case over voltage/temperature and across outputs.
- Bandwidth ambiguity: jitter numbers without integration bounds are not comparable.
- Test condition mismatch: output load, termination, and measurement point can change results dramatically.
- Single-output data: multi-output systems need channel-to-channel behavior, not only “one channel” plots.
Request a Quote
H2-13. FAQs (Troubleshooting) + JSON-LD
These FAQs close long-tail field issues without expanding the main text. Each answer uses a fixed 4-line structure: Likely cause / Quick check / Fix / Pass criteria.
Why does additive jitter look worse on the scope than in the datasheet?
Why do only some outputs fail while others pass on the same fanout buffer?
HCSL clock shows big overshoot—what is the first termination check?
Example materials (verify): 49.9Ω 0402 Yageo RC0402FR-0749R9L; 100Ω diff 0402 Yageo RC0402FR-07100RL; 0Ω option link Yageo RC0402JR-070RL.
LVPECL output looks “clipped” or shifted—what biasing is missing?
Example materials (verify): 49.9Ω 0402 Yageo RC0402FR-0749R9L; 0.1µF 0402 Murata GRM155R71C104KA88D (AC coupling/decoupling use-case dependent).
Why does enabling outputs (OE) cause a one-time glitch at endpoints?
How do I separate buffer additive jitter from board-induced jitter quickly?
Why does SSC work at the source but the endpoint PLL loses lock after fanout?
Channel-to-channel skew changes over temperature—what usually dominates first?
Why does adding a series resistor help SI but sometimes worsens jitter?
Probing the clock changes the measured jitter—how to measure without loading?
Example materials (verify): 100Ω diff termination Yageo RC0402FR-07100RL; 0.1µF Murata GRM155R71C104KA88D; ferrite bead Murata BLM18AG601SN1D for supply isolation (if needed by budget).
