Electronics Guide

Articles in This Category

Why Topology Shapes Signal Integrity

Every routing topology is, electrically, a network of transmission lines joined at branch points and ended by loads. Wherever the impedance seen by an advancing wave changes, part of that wave reflects. The reflection coefficient at a discontinuity is

Γ = (Z_L − Z₀) / (Z_L + Z₀)

where Z₀ is the characteristic impedance of the line the wave is traveling on and Z_L is the impedance it encounters. The transmitted amplitude is 1 + Γ. That single expression accounts for most of what distinguishes one topology from another, because topology is largely a matter of how many values of Z_L the signal meets on its way to each receiver, and when the resulting echoes arrive.

Two variables dominate. The first is the stub: any length of trace that branches off the main path and is not ended in its characteristic impedance. A stub loads the line capacitively at low frequency and resonates at high frequency, and the disturbance it launches grows with both its length and the sharpness of the edge. Topologies that avoid stubs, such as point-to-point and fly-by daisy chains, scale to the highest data rates. Topologies that depend on stubs, such as classic multi-drop buses, stay confined to lower speeds or to short electrical lengths.

The second variable is path balance. When one source must reach many loads, the difference in flight time between the longest and the shortest branch consumes timing margin directly. Distribution topologies are therefore judged largely by how equal their branches can be made, and by whether the receiving devices can measure and compensate for whatever imbalance remains.

When Topology Starts to Matter

Topology is a signal integrity concern only when the interconnect behaves as a transmission line rather than as a wire. The threshold depends on the edge rate, not on the clock frequency. A net must be treated as a transmission line once the round-trip delay along it becomes comparable to the rise time of the signal it carries, because only then does a reflection return late enough to disturb the edge that launched it rather than merging invisibly into it.

Propagation delay on a printed circuit board is roughly 6 to 7 nanoseconds per meter, or on the order of 150 to 180 picoseconds per inch, with outer-layer microstrip faster than stripline because part of its field travels in air rather than in the dielectric. A 1 nanosecond edge therefore tolerates several inches of unmanaged copper, while a 50 picosecond edge from a modern serializer turns a fraction of an inch into a transmission line. This is why topologies that were perfectly serviceable for a 33 megahertz parallel bus became unusable as edge rates sharpened, even on nets whose data rate had not changed.

The practical consequence is that topology decisions belong to the earliest stage of a design, before placement is fixed. Once the connectivity is committed—how many loads share a net, and roughly where they sit—the achievable signal integrity is largely determined, and layout can only realize or squander it.

The Fundamental Topologies

Point-to-Point

A single driver connects to a single receiver over one controlled-impedance line. With only two ends to manage, the path can be matched cleanly—series termination at the source, parallel termination at the receiver, or both—and no intermediate stubs launch reflections. Point-to-point is the topology of choice for the highest-speed signals, and it underlies essentially every modern multi-gigabit serial link.

Its limitation is fan-out. Reaching additional loads requires additional drivers, additional links, or a switch, and each of those costs pins, power, and area. Serial link design accepts that cost deliberately: rather than sharing one conductor among many devices at modest speed, it dedicates a clean pair to each connection and runs it far faster. The trade is almost always favorable above about a gigabit per second, because the cost of a driver falls with process scaling while the cost of a mismatched net does not.

Multi-Drop

Several receivers connect to a common line, each through a short branch, so a single driver broadcasts to all of them. The shared conductor is economical in pins and copper, and it was the basis of the parallel buses that once dominated processor and peripheral interconnect.

The cost is electrical. Each tap adds device and package capacitance plus a stub, the distributed loading lowers the effective impedance of the line and slows propagation along it, and reflections originate at every tap rather than at one place a designer can terminate. As edge rates sharpen, these accumulated echoes close the timing window from both sides. Multi-drop survives today mainly where its simplicity is worth more than its bandwidth: control, configuration, and sensor buses running at kilobit to low-megabit rates over short distances, and differential linear buses in which the whole cable behaves as one continuous, doubly terminated line.

Daisy Chain and Fly-By

Loads are connected in series, the line entering and leaving each device in turn so that one trace threads through every load along its length. Because the through path stays continuous, the stubs that plague multi-drop shrink to the short connection inside each package, and the routing is compact.

The signature cost is cumulative propagation delay. A signal reaches the last device measurably later than the first, so the skew along the chain must either fit inside the timing budget or be actively compensated. The most important instance is the fly-by topology introduced with DDR3 for the command, address, control, and clock signals: the line passes each device in sequence and terminates once beyond the last, with the terminating resistors placed on the module itself. Fly-by replaced the branched routing of earlier generations precisely because removing the stubs allowed the interface to reach far higher data rates, and the deliberate flight-time skew it creates is absorbed by training rather than by layout.

Tree and Star

A common source fans out through one or more branch points to reach many loads over paths designed to be as equal as possible. In a star, the branches radiate from a single node. In a tree, they split through a hierarchy of junctions, frequently with an active buffer at each level to restore drive and to isolate the branches from one another. The symmetric H-tree, in which every split is a matched pair, is the limiting case: it makes the electrical length from source to every leaf identical by construction.

Balanced branching is the standard way to distribute a clock or reference across a board, because matched branch lengths hold the skew between loads small. The difficulty is the junction itself. Power divides among the legs, the impedance looking into the junction drops, and each leg generally needs its own termination, which makes both the layout and the resistor budget demanding.

Hybrid Arrangements

Real interfaces rarely use one topology for every signal. A memory channel routes its data groups point-to-point or as short chains while routing command and address fly-by. A JTAG chain is a hybrid by definition: TDI and TDO thread serially through the devices, forming one long shift register, while TCK and TMS are broadcast to every test access port so that all of them change state together. Recognizing which signals on a bus belong to which topology—and terminating each accordingly—matters more than labeling the interface as a whole.

Stubs, Loading, and Branch Points

The Stub

A short spur off the main line behaves as a lumped capacitance and merely rounds the edge. Once its round-trip delay approaches the rise time, it reflects a distinct wave that arrives after the edge and lands inside the data window. At a quarter wavelength it resonates and can null the signal outright. The useful design rule follows directly: keep the round-trip delay of every stub small compared with the fastest edge the net carries, and remember that this budget tightens every time a driver generation gets faster, even if the data rate does not change.

Stubs hide in more places than the schematic suggests. The unused barrel of a through-hole via is the most common, which is why high-speed designs back-drill the unused portion away or use blind and buried vias. Test points, unpopulated footprints, connector fingers, series resistors placed away from the pin they serve, and probe pads all add copper that no one intended to be part of the net.

Distributed Loading

When many loads share a line, their capacitance adds to the intrinsic capacitance of the transmission line and changes its properties. If C₀ is the intrinsic capacitance of the line and C_L is the added load capacitance distributed along it, the loaded impedance falls and the propagation delay rises by the same factor, the square root of (1 + C_L/C₀).

Two consequences follow. First, a bus designed and fabricated for 50 ohms unloaded may present appreciably less than that once populated, so the terminating resistor chosen from the fabrication drawing is wrong. Careful multi-drop design either accounts for the loaded value when selecting termination or raises the unloaded impedance deliberately so that the populated bus lands near the target. Second, the loading is only uniform if the loads are evenly spaced; clustered devices create a locally slow, locally low-impedance region that reflects at both of its ends.

The Branch Point

Where a line splits into two branches of equal impedance, the incident wave sees the two branches in parallel. A 50-ohm line meeting two 50-ohm branches therefore encounters 25 ohms, and the reflection coefficient is −1/3: about a third of the wave returns toward the source inverted, and about two-thirds continues into each branch. This is why an unbuffered split produces a visible step on the waveform even when everything is nominally matched.

Designers respond in three ways. They can balance the branch lengths so that the reflections from the far ends return together and the receivers see a common, predictable disturbance rather than a scattered series of them. They can raise the impedance of the branches so that their parallel combination approaches the impedance of the feeding line, at the cost of narrower traces and higher loss. Or they can place an active buffer at the junction so that each branch becomes an independent point-to-point net driven from its own low-impedance source. Buffering costs power, area, and a small amount of added jitter, and it is the usual answer whenever the fan-out is large or the skew budget is tight.

Termination by Topology

Termination follows the topology rather than the reverse, and the number of matched ends a topology offers determines which schemes are even available.

Point-to-point links accept the full range of methods because they have exactly one far end. Series termination at the source, sized so that the driver output impedance plus the resistor equals the line impedance, absorbs the returning reflection and dissipates no static power, which suits single-load nets and clock lines well. Parallel termination at the receiver holds the line matched continuously and preserves the edge best, at the cost of steady current. Thevenin and AC-coupled variants trade static dissipation against complexity and component count.

Daisy chains are terminated at the electrical end of the chain, so the final device sees a matched line and the wave dies there instead of returning through every device it already passed. Placement is unforgiving: a terminator beyond the last load is correct, while a terminator before it leaves an unterminated tail.

Tree and star topologies generally need one terminator per leg, because each leg is its own line with its own far end. Series termination at the branch point is common for buffered clock distribution, since one resistor per output serves both to match the leg and to slow the edge slightly.

Multi-drop is the hardest case, because reflections originate at every tap and no single resistor absorbs them all. The classic remedy is a linear bus with no meaningful stubs and a terminator at each physical end, which is how CAN and RS-485 segments are built, typically with 120 ohms at each end to match the cable. Where the loads must genuinely tap a shared trace, the designer minimizes stub length, spaces the loads evenly, and accepts a bandwidth ceiling.

Across all of these, on-die termination integrated into the driver or receiver has become the preferred mechanism at high speed. Placing the matching element inside the silicon suppresses the package and stub reflections that an external resistor, sitting a few millimeters away on the board, can never reach. On-die termination also allows the value to be calibrated against a reference resistor and switched by direction, which a fixed external network cannot do.

Skew, Timing, and Training

Every topology that reaches more than one load creates skew, and the design either removes it geometrically or measures and cancels it electrically.

Geometric removal means length matching: serpentine sections, deliberate detours, and symmetric branching that equalize flight time across loads. It is the traditional approach and remains the right one for clock trees, for the lanes within a parallel group, and for the two halves of a differential pair. It costs board area, it consumes routing channels, and it has limits, because serpentine sections that are packed too tightly couple to themselves and lower the local impedance.

Electrical cancellation means letting the topology be as skewed as routing convenience demands and then training the receiver to it. Fly-by memory routing is the definitive example. Because the clock arrives at each device at a different time, DDR3 introduced write leveling: during initialization the controller delays each data strobe until it aligns with the clock as that clock arrives at the device the strobe serves, and a complementary adjustment handles the read direction. Later generations extended the same philosophy. DDR5 carries the fly-by discipline forward and additionally divides each module into two independent sub-channels, each with its own command and address bus, so that fewer devices hang on any one line; JEDEC speed grades for DDR5 begin at 3,200 megatransfers per second, and the JESD79-5C revision published in 2024 raised the defined maximum to 8,800, well beyond anything a stubbed topology could support.

Training changed the economics of topology selection. Once a link can measure and compensate its own skew, a topology that routes compactly but unevenly becomes preferable to one that routes symmetrically but at the cost of stubs, because skew is correctable and reflections are not.

Topology in Practice

Serial Links

PCI Express, USB, SATA, and Ethernet all use point-to-point differential pairs, with switches or hubs providing fan-out. The topology is chosen so that each pair sees exactly one transmitter and one receiver, leaving equalization and clock recovery to deal with loss and jitter rather than with reflections. Where several devices must be reached, the architecture adds a switch rather than a tap.

Control and Sensor Buses

I2C, SMBus, and similar single-ended buses are multi-drop by design and are limited chiefly by total bus capacitance, which the specifications cap explicitly. The I2C specification allows at most 400 picofarads on each bus line in Standard-mode and Fast-mode, 550 picofarads in Fast-mode Plus, and 100 picofarads in High-speed mode for the full 3.4 megabits per second, a limit that relaxes to 400 picofarads if the rate is held to 1.7 megabits per second. Every device, connector, and inch of trace spends part of that budget, which is why a long or heavily populated bus must either slow down or add an active buffer. CAN and RS-485 are also multi-drop, but they run as differential linear buses with a terminator at each physical end and short device stubs, which is what allows them to span tens of meters at rates single-ended multi-drop could not sustain.

Memory Interfaces

Modern DRAM channels mix topologies within one interface. Command, address, control, and clock signals travel fly-by, with the terminating resistors on the module. Data and strobe signals are grouped into byte lanes routed point-to-point or in short chains to the devices that own them, with on-die termination at whichever end is receiving. The split exists because command and address must reach every device while data need reach only the device it belongs to.

Clock and Reference Distribution

Clock networks are the natural home of trees and stars. A buffered tree with matched branch lengths, or a symmetric H-tree, distributes an edge to many loads with small skew, and phase-locked zero-delay buffers can remove the insertion delay of the distribution itself. Because a clock is periodic and every load samples the same edge, skew and jitter matter far more here than the small amplitude losses a designer would worry about on a data line.

Test and Configuration Chains

Boundary scan, device configuration, and many display and lighting interfaces use daisy chains because the topology matches the data flow: information passes through each device in order. As the JTAG example shows, these chains almost always pair a serial data path with a broadcast clock and control path, and the broadcast signals are the ones that need topology and termination attention, since they must arrive at every device at once.

Choosing a Topology

Topology selection follows from the number and placement of loads, the data rate and edge rate, the skew budget, and the available board area and power. A single fast load points to point-to-point with source or end termination. Many loads that must be reached compactly at high speed point to a fly-by daisy chain with end termination and a training scheme to absorb the skew. A clock or reference bound for many loads with a tight skew budget points to a balanced tree or star with per-branch termination, usually buffered. A handful of slower loads on a shared control bus is still served well by a simple multi-drop arrangement, and a long cable run of such loads by a doubly terminated differential linear bus.

The choice is worth testing before the layout exists. Pre-layout simulation with driver and receiver models, exercising the candidate topologies over the expected range of trace lengths and load counts, answers the questions that matter—how much eye closes per additional load, how long a stub the interface tolerates, where the terminator must sit—while changing the answer is still free. After placement, the same simulation flow run on extracted geometry verifies that the built net behaves as the chosen topology predicted.

Bringing It Together

No single topology is best; each trades fan-out, speed, skew, and routability against the others. Point-to-point maximizes signal integrity at the expense of fan-out. Multi-drop maximizes sharing at the expense of speed. Daisy chain and fly-by reach many loads compactly while incurring chain delay that training must absorb. Tree and star deliver balanced low-skew distribution at the cost of layout effort and buffering. Sound high-speed design begins by matching the topology to the connectivity the interface actually requires, then controls the impedance of every segment and terminates each path to suppress the reflections the chosen geometry creates.

The topics in this category examine each strategy in depth, from the clean simplicity of point-to-point links to the stub management of multi-drop buses, the chain delay and write leveling of daisy-chain routing, and the balanced fan-out of tree and star distribution.

Related Topics