Electronics Guide

Articles in This Category

Why the Interconnect Matters at High Speed

A signal does not perceive a bond wire, a solder ball, or a connector contact as a discrete object; it perceives a change in the surrounding electromagnetic environment. Wherever the cross section of the conductor and its return path changes, the instantaneous impedance changes, and part of the signal reflects. What makes these structures problematic is the bandwidth of modern edges. The frequency content of a digital signal extends well beyond its bit rate. A common rule of thumb places the significant bandwidth of an edge near 0.35 divided by the 10 to 90 percent rise time, and the related knee-frequency convention places it near 0.5 divided by the rise time. Under either convention a 30-picosecond edge carries meaningful energy past 10 gigahertz, and at those frequencies a few millimeters of lead, a stub on a connector pin, or an unterminated via behaves as a distributed element rather than a lumped one.

A useful test of whether a structure may still be treated as a lumped parasitic is to compare its propagation delay with the rise time: as a working rule, the structure behaves as a lumped element only when its delay is smaller than roughly one-sixth of the rise time. Signals travel through typical package and board dielectrics at something near 6 to 7 picoseconds per millimeter. For a 30-picosecond edge, therefore, the lumped region ends at well under a millimeter of physical length. Almost every feature in a package or a connector is longer than that, which is why modern practice models these structures as transmission lines and extracts them with three-dimensional field solvers rather than reducing them to a single inductor and capacitor.

Two consequences follow. First, a discontinuity that is electrically short at 1 gigahertz can be a meaningful fraction of a wavelength at 50 gigahertz, so it reflects more strongly and resonates where it once merely attenuated. An unused length of plated through-hole below a connector pin, for example, forms a stub that presents a quarter-wave resonance and can notch the channel response squarely inside the signal band; backdrilling that stub away is now routine practice. Second, the same parasitics that distort signals also corrupt the power delivery network. Inductance in the package and on the board limits how quickly charge can reach switching transistors, producing the supply droop and ground bounce that erode timing margins. The package and interconnect thus sit at the intersection of signal integrity and power integrity, and a sound design addresses both at once.

The Parasitic Budget

Each element in the path contributes a characteristic mix of inductance, capacitance, and resistance, and good design begins with knowing which parasitic dominates where.

• Inductance. Long, thin conductors with a distant return path are inductive. A wire bond contributes roughly 1 nanohenry per millimeter, a rule of thumb that makes a one- to two-millimeter bond a serious limitation above a few gigahertz. Series inductance slows edges, rings against nearby capacitance, and is the principal source of simultaneous-switching noise on power and ground. Because inductance scales with the area of the loop formed by the signal and its return, the cure is almost always geometric: shorten the conductor, or bring the return closer.
• Capacitance. Broad conductors close to a plane, such as a die pad, a ball-grid-array land, or a connector footprint, are capacitive. Shunt capacitance loads the driver, rounds edges, and lowers the local impedance, creating a dip that reflects fast transitions. A capacitive pad is often relieved by clearing the reference plane beneath it, which trades a little added inductance for a large reduction in capacitance and moves the net impedance back toward the target.
• Resistance. Fine traces, micro-bumps, and contact interfaces add series resistance that attenuates the signal and rises further at high frequency as the skin effect confines current to the conductor surface. Skin depth in copper falls below a micrometer around 10 gigahertz, so only a thin shell of each conductor carries current and the effective resistance grows roughly with the square root of frequency. Surface roughness, added deliberately to make copper adhere to laminate, compounds the loss. In connectors and sockets the contact resistance is also a reliability parameter, sensitive to normal force, plating, and fretting.

Mutual parasitics matter as much as self parasitics. Adjacent bond wires, neighboring package balls, and parallel connector pins couple both inductively and capacitively, and that coupling appears at the far end as crosstalk. In a dense connector field the aggressor is rarely a single neighbor; the victim pair sees the summed contribution of every surrounding pin, which is why high-speed connector designers spend as much effort on ground pin placement and internal shielding as on the impedance of the signal contacts themselves.

Designers manage this budget by keeping interconnect lengths short, by pairing every signal with a close and continuous return, and by matching impedance across each transition so that the chain looks as nearly as possible like one uniform transmission line. Where a discontinuity is unavoidable, it is compensated: a capacitive land may be tuned with an adjacent inductive section, and a connector launch is engineered so its impedance tracks the board it joins. The goal is not to eliminate every parasitic, which is impossible, but to keep the reflections small, to place any residual resonance outside the band of interest, and to leave the remainder to the equalization in the transceiver.

The Path from Die to Board

Following a signal outward from the silicon shows how the topics in this category fit together. Each stage exists to bridge a difference in scale, and each bridge exacts an electrical price.

Off the Die: Wire Bonds and Flip-Chip

At the die, fine aluminum or copper pads on a pitch measured in micrometers must connect to the far coarser world of the package. Two approaches dominate. Wire bonding runs a fine gold or copper wire, a few tens of micrometers in diameter, from each pad to a lead frame or substrate. It is mature, inexpensive, and well suited to peripheral pad arrangements, but the bond loop is inductive and the pads are confined to the die edge, which limits both the count and the speed of the connections. Flip-chip mounting instead turns the die over and joins its pads directly to the substrate through an area array of solder bumps. The path shortens to a fraction of a millimeter, connections distribute across the whole die face rather than crowding its perimeter, and power can be delivered near the circuits that draw it. The IC package effects topic treats these structures and their parasitics in detail.

Bridging the Pitch Gap: Interposers and Bridges

Where a single substrate cannot bridge the pitch gap, or where several dies must sit close together and exchange enormous bandwidth, an interposer is inserted between the chips and the package substrate. Silicon interposers carry dense redistribution wiring and through-silicon vias, and they connect to the dies above through micro-bumps whose pitch in current production sits in the range of roughly 40 to 55 micrometers. That pitch has a practical floor near 25 micrometers, below which the solder volume becomes too small to form a reliable joint. Finer connection therefore requires a different mechanism: hybrid bonding, in which copper pads bond directly to copper and the surrounding dielectric bonds oxide to oxide, reaching pitches below 10 micrometers with no solder at all. Organic interposers and redistribution-layer fan-out structures offer lower-cost alternatives at coarser pitch, and embedded silicon bridges place a small, dense die inside an otherwise organic substrate to provide high-density wiring only where two neighbors must talk. These structures are the foundation of 2.5D and 3D integration, covered under interposer and substrate design.

The Substrate Itself

The package substrate fans the array out from micrometer pitch to the millimeter-scale pitch a board can accept, and its material properties set much of the achievable performance. Organic buildup laminates dominate: they are inexpensive and well understood, but they warp under thermal load, and warpage at large body sizes limits both assembly yield and the fine lithography that dense wiring requires. Loss is the other constraint, since the dielectric the wiring sits in determines how much of a 50 gigahertz signal survives the journey across a large package. Glass core substrates, valued for flatness, dimensional stability, and the ability to support very large body sizes, have moved from research into pilot lines and first products, though yield remains the gating factor in their adoption.

Onto the Board

The completed package then reaches the board through a soldered ball or land array, or through a separable socket, and travels between boards through connectors and cables. Where the assembly must bend or fold into a three-dimensional envelope, the path continues across flexible and rigid-flex circuits, which trade some impedance control and shielding for mechanical freedom. Each handoff is an opportunity to preserve or to spoil the signal.

Sockets, Connectors, and the Mating Interface

Separable interconnects add a constraint the rest of the path lacks: a mechanical mating interface that must conduct reliably over many cycles while staying electrically transparent. A socket lets a processor be replaced or upgraded; a backplane connector lets line cards be inserted into a chassis. Both introduce a contact pair whose resistance, inductance, and capacitance must be controlled, and whose geometry tends to create stubs and impedance steps that reflect high-speed energy.

Modern high-speed connectors meet these demands with impedance-controlled geometries, careful ground referencing, and internal shielding that suppresses coupling between adjacent pairs. High-volume backplane and cabled interconnect now runs at 112 gigabits per second per lane, which uses four-level pulse-amplitude modulation at 56 gigabaud and so places its Nyquist frequency near 28 gigahertz. The leading edge has moved to 224 gigabits per second per lane: the same PAM4 format at 112 gigabaud, with a Nyquist frequency near 56 gigahertz, so the connector and its launch must hold impedance and suppress crosstalk across a band extending well beyond 50 gigahertz. The IEEE 802.3dj task force and the Optical Internetworking Forum's CEI-224G project define the electrical interfaces at this rate.

The mechanical side is equally demanding. Contact resistance depends on normal force, on the wipe that scrubs oxide from the mating surfaces during insertion, and on the plating that protects them. Gold plating resists oxidation and suits many mating cycles at a higher cost; tin is cheaper but is vulnerable to fretting corrosion, in which small relative motions under vibration or thermal cycling gradually build an insulating layer of oxide debris at the interface. Ratings differ accordingly: a consumer or server socket may be specified for only a few tens of insertions, while a test or burn-in socket must survive far more. Press-fit compliant pins press into plated through-holes to form a gas-tight, solderless joint, avoiding the thermal stress of soldering a large connector and providing a repeatable contact; the general requirements for such press-in connections are set out in IEC 60352-5.

These constraints trade against one another. Higher normal force lowers and stabilizes contact resistance but raises insertion force, which matters when a connector carries hundreds of contacts. Adding ground pins improves isolation and lowers return inductance but consumes the pin count available for signals. Wider contacts carry more current but add capacitance. Resolving these tensions is the subject of the socket and connector design topic.

Power Delivery Through the Package

The package is also the last stage of the power delivery network, and it is often the limiting one. A large processor may draw hundreds of amperes and change that draw by a substantial fraction within a few clock cycles. The resulting current transient develops a voltage across every inductance in the path, and the package contributes a significant share of that inductance between the board-level capacitors and the die.

The consequence is a division of labor across frequency. Bulk capacitors on the board handle slow transients. Ceramic capacitors near the package handle the middle band. Above roughly 100 megahertz the loop inductance through the package makes board-mounted capacitors ineffective, so the highest-frequency charge must come from capacitance placed on the package substrate, from deep-trench or metal-insulator-metal capacitors integrated into the die or interposer, and ultimately from the intrinsic capacitance of the on-die power grid itself. Some designs push further and place voltage regulation inside the package, shortening the high-current path dramatically at the cost of added thermal density.

Two failure modes recur. The first is simultaneous-switching noise, in which many outputs transition together and their aggregate current, flowing through shared power and ground inductance, momentarily shifts the reference seen by the remaining circuits. The second is anti-resonance: the inductance of one stage resonates against the capacitance of the next, producing an impedance peak at which a modest current transient generates an outsized voltage excursion. Both are properties of the package and its attachment, not of the silicon alone, which is why power integrity analysis must include an extracted package model rather than an idealized supply.

Reliability and the Mechanical Dimension

Electrical performance is bounded by what the assembly can survive. Silicon expands at roughly 3 parts per million per kelvin, while organic substrates, copper, and board laminate expand several times faster. Every thermal cycle therefore strains the joints between them. Underfill, an epoxy drawn beneath a flip-chip die, couples die and substrate so the strain is shared rather than concentrated in the solder, and it is essential at fine bump pitch.

The same stress reaches inward. Low-permittivity interlayer dielectrics improve on-die speed but are mechanically fragile, and the force transmitted through a bump during assembly or thermal cycling can crack them near the die corners, where the mismatch is greatest. This coupling between package mechanics and device behavior is the domain of silicon-package interaction. Warpage adds a further constraint: a substrate that bows during reflow may leave joints open at the die corners or bridge them at the center, and the tolerance shrinks as body sizes grow. These are not side concerns for the signal-integrity engineer, because an intermittent or cracked joint presents itself first as a marginal channel and only later as an outright failure.

Modeling, Extraction, and Measurement

Because package and interconnect structures are three-dimensional and electrically long, they resist hand analysis. Practice therefore rests on extraction and measurement. A field solver computes the electromagnetic behavior of a bump field, a via transition, or a connector footprint and reduces it to a scattering-parameter model, conventionally exchanged in Touchstone format. Those models cascade with board models and with behavioral transceiver models to predict the eye at the receiver. Buffer behavior is commonly described in IBIS models, with the algorithmic extension covering the equalization and clock recovery that modern serial links depend on.

A model is only as good as its passivity, causality, and reciprocity, and an extracted block that violates any of them will produce simulation results that look plausible and are wrong. Checking those properties before use is standard discipline. Measurement provides the reality check: a vector network analyzer characterizes the structure in the frequency domain, and time-domain reflectometry locates discontinuities along the path by showing where the impedance departs from its target. Both require that the fixture used to reach the device be removed from the result, and de-embedding methods for that purpose, including the fixture design and removal practices standardized in IEEE Std 370, have become as important as the measurement itself at these frequencies.

About This Category

The topics gathered here trace the interconnect from the silicon outward. IC package effects characterize the first transition off the die; interposer and substrate design bridge fine die pitch to the package and enable heterogeneous integration; socket and connector design govern the separable interfaces that join packages to boards and boards to one another; and flex and rigid-flex design extend the path into folded, three-dimensional assemblies. Read together, they provide the framework for treating the package and interconnect as a deliberate part of the high-speed channel, designed for signal integrity and power integrity rather than left to chance. The practical lesson is consistent across all four: the interconnect is not the passive plumbing between the interesting parts of a system, but a designed element whose parasitics, impedance profile, and mechanical margins decide how fast that system can run.

Related Topics