Electronics Guide

Stack Topologies: 2.5D, 3D, and Hybrid Arrangements

Two arrangements dominate multi-die packaging, and the vocabulary distinguishing them matters because their electrical characteristics differ substantially. In a 2.5D arrangement, dies sit side by side on a shared routing substrate—a silicon interposer, an embedded silicon bridge, or a fine-line redistribution layer—and communicate laterally through that substrate. No die sits on top of another, so no die must carry through-silicon vias for the sake of its neighbor. In a true 3D arrangement, dies stack vertically and signals pass through the body of the silicon itself.

The 2.5D approach trades interconnect length for simplicity. Lateral routes across an interposer run millimeters rather than micrometers, so they behave as short transmission lines with measurable loss and require attention to impedance and return path. In exchange, the thermal problem is far easier, because every die has direct access to the package lid, and the assembly avoids thinning dies to the point where they become fragile. Most large accelerator packages that pair a compute die with high-bandwidth memory stacks use exactly this arrangement at the package level.

True 3D stacking gives the shortest possible connection and the highest connection density, but every die below the top of the stack must be thinned and must carry vertical vias, and every die except the outermost one is thermally buried. Real products routinely combine both: a 2.5D package hosting a logic die alongside memory stacks that are themselves 3D assemblies. The architecture question is therefore not which approach to adopt but where in the system each boundary should fall.

Face-to-Face and Face-to-Back Orientation

Every die has a face, the side carrying transistors and the metal interconnect stack, and a back, the bulk silicon behind it. The relative orientation of two bonded dies is the single most consequential architectural decision in a stack, because it determines whether a signal crossing between them must pass through silicon at all.

In a face-to-face bond, two dies are flipped toward each other and their topmost metal layers are joined directly. The connection is the shortest available: metal to metal, with no via passing through bulk silicon. Parasitic capacitance and resistance are minimal, and the arrangement supports the finest connection pitches. Its limitation is arithmetic. Only two dies can be joined face to face, so a face-to-face pair still needs through-silicon vias in one of the two dies to reach the package below, and the pair cannot be extended into a taller stack without adopting a different scheme for the additional tiers.

In a face-to-back bond, one die's face is joined to the back of the die above it. This orientation extends indefinitely, which is why memory stacks and other tall assemblies use it. The cost is that signals must traverse a through-silicon via in every die they pass, and each via contributes series resistance and inductance plus capacitance to the surrounding substrate. Because the substrate is a semiconductor rather than an insulator, that capacitance varies with bias as a depletion region forms and collapses, and adjacent vias couple through the shared silicon. A face-to-back link is therefore electrically busier than a face-to-face one, and its behavior depends on via geometry, spacing, and the presence of grounded guard vias.

Back-to-back arrangements exist but are uncommon, since they place two thicknesses of bulk silicon between the active layers and give up most of the proximity advantage. The practical design space is the choice of where to spend the one available face-to-face interface, and the answer usually places it at the highest-bandwidth, most latency-critical boundary in the system.

Die-to-Die, Die-to-Wafer, and Wafer-to-Wafer Bonding

Independent of orientation, a stack can be assembled at three granularities, and the choice shapes both yield economics and achievable pitch.

Wafer-to-wafer bonding joins two full wafers before singulation. It achieves the finest alignment and the highest throughput, because a single alignment step serves every die on the wafer simultaneously. Its weakness is yield: the two wafers must have identical die sizes and identical step-and-repeat patterns, and a defective die on either wafer is paired with a good die on the other, destroying both. Wafer-to-wafer bonding is therefore attractive when both layers are small, mature, and high yielding—stacked image sensors are the classic example—and unattractive when either layer is large or new.

Die-to-wafer bonding places individually tested dies onto a full base wafer. Only dies that passed test are placed, so a defect on one layer no longer condemns its partner, and the two layers need not share a die size. The cost is throughput, since each die requires its own pick, align, and place operation, and alignment accuracy is harder to hold than in the wafer-to-wafer case. Most high-value 3D stacking today uses this approach.

Die-to-die bonding assembles singulated dies onto singulated dies, offering the greatest flexibility in mixing die sizes and sources at the lowest throughput. It suits low-volume, high-mix, or heterogeneous assemblies where the parts come from different fabs entirely.

Micro-Bump Signal Integrity

Micro-bumps are the established electrical connection between vertically adjacent dies, or between a die and an interposer. They are miniaturized solder joints, typically formed on a copper pillar with a thin solder cap, and current production geometries place them roughly ten to twenty-five micrometers across on pitches of roughly twenty-five to fifty-five micrometers. Earlier high-bandwidth memory generations sat near the upper end of that band; HBM3E has tightened to roughly twenty-five micrometers, and HBM4 pushes finer still, which is also where advanced-package die-to-die interfaces operate. The joint itself stands only on the order of ten to thirty micrometers tall, so absolute resistance and inductance per connection are small.

Small does not mean negligible. The inductance that matters is not that of a single bump but the loop inductance formed by a signal bump and its return path through the nearest ground bumps. A signal bump surrounded by ground bumps presents a small loop and a well-behaved impedance; the same bump placed at the edge of an array with its nearest ground several pitches away presents a much larger loop and a correspondingly worse transition. Bump map assignment is consequently a signal integrity task, not a mechanical one, and the ratio and placement of ground bumps within the array set the achievable data rate as firmly as any driver characteristic.

Capacitive coupling between adjacent bumps produces crosstalk that limits dense arrays. The short spacing and the tall aspect of a pillar create fringing capacitance to every neighbor, and the underfill that mechanically supports the joint raises the effective permittivity of the gap. Designers manage this by interleaving grounds, by assigning aggressive and sensitive nets to non-adjacent positions, and by extracting the whole array with a three-dimensional field solver rather than a lumped estimate. The extraction must include the redistribution layers on both dies, because the transition from on-die metal through the bump and into the receiving die is a continuous structure whose behavior no single element captures.

Manufacturing variation adds a statistical dimension. Bump height varies across a die, placement accuracy has a tolerance, and the two mating surfaces are never perfectly coplanar. Unequal joint geometry produces unequal current sharing across nominally parallel connections, which matters most for power delivery, where a bump carrying more than its share heats faster and ages faster. Robust designs add redundant power connections, hold current density well below the nominal limit, and verify the interface against the statistical distribution of bump parameters rather than a single nominal case.

Direct Copper Hybrid Bonding

Hybrid bonding removes the solder joint entirely. Copper pads embedded in a dielectric are polished to an extremely flat, clean surface on both dies; the dielectric surfaces are brought into contact and bond at room temperature, and a subsequent anneal expands the copper pads into direct metal-to-metal contact. The result is a bumpless interface in which the two metal stacks are effectively continuous. The name reflects the two simultaneous bonds, dielectric to dielectric and metal to metal.

The electrical payoff is large. Removing the solder ball and its underfill eliminates the dominant parasitic capacitance and resistance of the interface, shortens the electrical path to well under a micrometer, and lowers the energy spent moving each bit across the boundary. The metal-to-metal contact also conducts heat considerably better than a solder joint surrounded by polymer underfill, which relieves one of the worst thermal bottlenecks in a tall stack. Because the connection no longer requires room for solder to reflow without bridging, pitch scales far beyond what micro-bumps permit.

Pitch is where the two technologies separate most sharply. Hybrid bonding reached roughly nine micrometers in high-volume manufacturing, industry roadmaps target the three-to-four-micrometer range and below, and laboratory demonstrations have gone to sub-micrometer pitch. Against a micro-bump floor in the tens of micrometers, this represents an increase in connection density of one to two orders of magnitude for the same area, which changes what an interface can be: at fine enough pitch, a die-to-die boundary starts to resemble on-die wiring rather than a package-level link, and functional blocks can be partitioned across the boundary rather than merely communicating across it.

The obstacles are process obstacles. Bonding surfaces must be flat to within nanometers and free of particles, because a single particle prevents contact over a region far larger than itself. Copper pad recess must be controlled precisely, since too little expansion leaves an open and too much causes dishing or extrusion. Alignment tolerance tightens in proportion to pitch. These requirements explain why adoption has been selective rather than universal: high-bandwidth memory, for instance, has continued to use solder micro-bumps through the HBM4 generation, with hybrid bonding held in reserve for later stacks where die thinning and stack height leave no alternative.

Die-to-Die Interface Design and Standards

A die-to-die link exploits the fact that its channel is thousands of times shorter than a board-level one. Instead of a few differential lanes running at very high rate with heavy equalization, a die-to-die interface uses a wide, slow, single-ended parallel bus: hundreds to thousands of lanes, low voltage swing, simple drivers, and a forwarded clock rather than embedded clock recovery. The channel is short enough that reflections settle within a fraction of a bit period and the loss slope over the band of interest is small, so continuous-time equalization and decision-feedback equalization can often be omitted altogether. Removing them is precisely where the energy saving comes from.

The industry has converged on shared specifications for these links so that dies from different vendors can be assembled together. The Universal Chiplet Interconnect Express (UCIe) specification defines a die-to-die physical layer, link layer, and protocol mapping, and it distinguishes package profiles because the achievable pitch determines the achievable design. A standard-package profile targets conventional organic substrates at bump pitches on the order of one hundred micrometers; an advanced-package profile assumes a silicon interposer or embedded bridge at pitches of roughly twenty-five to fifty-five micrometers; and a 3D profile added in later revisions targets hybrid-bonded stacks, spanning pitches from roughly ten to twenty-five micrometers down to one micrometer or less. UCIe 3.0, published in August 2025, raised the per-lane signaling rate to 48 and 64 GT/s, up from 32 GT/s in earlier revisions. Alongside UCIe, the Open Compute Project's Bunch of Wires specification and the Advanced Interface Bus contributed to CHIPS Alliance offer openly published alternatives, and many vendors continue to run proprietary links tuned to their own packaging.

Wide parallel links change what reliability means. With thousands of lanes, the probability that every lane is perfect is low, so interfaces include spare lanes and a repair mechanism that maps a failed lane onto a spare during link training. Error handling generally relies on a cyclic redundancy check with link-level retry rather than continuous forward error correction, because the raw error rate is low and retry costs nothing when errors are rare; forward error correction appears at the highest signaling rates, where the latency cost of retry begins to outweigh its simplicity. Published energy targets for advanced-package links sit well below one picojoule per transferred bit, an order of magnitude or more better than a SerDes channel crossing a printed circuit board.

Power management deserves architectural attention because a wide bus idles expensively if left fully active. Interfaces therefore support dynamic width reduction and low-power link states, trading latency on wake for static power while idle. The control logic must coordinate these transitions across both dies without losing training state, which is one of the reasons a die-to-die link, despite its electrically simple channel, carries a substantial digital protocol layer.

Interposer and Bridge Routing Strategies

An interposer is a routing substrate that fans the fine-pitch connections of a die out to the coarser pitch of a package, and that carries lateral connections between dies mounted on it. Silicon interposers use semiconductor fabrication to reach the finest line widths and via densities, letting them carry the thousands of parallel connections a memory stack requires while holding controlled impedance. Their limitation is cost and reticle size, since an interposer large enough to host several dies approaches or exceeds the maximum field a lithography scanner can print in one exposure, requiring stitching or a different approach.

Embedded silicon bridges address that limitation by placing a small piece of fine-pitch silicon only where two dies must communicate, embedding it in an otherwise conventional organic substrate. The dense routing exists only where it is needed, the package can grow without a correspondingly large silicon layer, and cost falls. The electrical trade is a discontinuity where signals leave the bridge for the organic substrate, and a more complex assembly process. Organic and fine-line redistribution interposers occupy a middle position, offering lower loss than silicon at a given geometry, because organic dielectrics are less lossy than a conductive silicon substrate, but coarser achievable pitch.

Routing within any interposer balances signal integrity against power delivery and layer count. High-speed routes need a continuous reference plane beneath them, tight coupling within differential pairs, and separation from aggressors. Power distribution needs wide, low-impedance metal and dense via arrays. These compete for the same small number of metal layers, and the resolution is usually a deliberate layer assignment: dedicated plane layers for power and ground, signal layers sandwiched between them, and a routing plan that keeps every high-speed net adjacent to an unbroken reference.

Thermal behavior also influences interposer design. Silicon conducts heat far better than organic laminate, which helps spread heat laterally away from a hot die, but the multi-layer metal and dielectric stack above the silicon impedes vertical flow. Practical designs add thermal vias and copper fill in otherwise unused regions to improve conduction, and place high-current connections where they contribute usefully to heat spreading. More aggressive concepts, including microfluidic channels etched into the interposer itself, have been demonstrated in research but remain outside volume production.

Power Delivery Through TSVs

Current reaching a die in the middle of a stack must climb through the through-silicon vias of every tier beneath it. Unlike signal vias, which carry brief transient currents, power vias sustain high continuous current, and their resistance and inductance appear directly as supply droop and ground bounce at the point where they are hardest to compensate.

The architecture sets the terms of the problem. Die thinning fixes the length of every power via, tier count fixes how many series interfaces a load on an upper die sits behind, and the bonding scheme fixes what the current must cross at each boundary. That last term usually binds. Electroplated copper interconnect tolerates current densities on the order of a million amperes per square centimeter, whereas solder micro-bumps begin to show electromigration damage near ten thousand amperes per square centimeter—roughly two orders of magnitude lower—so the joint, not the via, defines the design limit at bump pitches below one hundred micrometers. Designs respond with redundant power connections that degrade gracefully if an individual path opens, and with verification that the supply still regulates after a specified fraction of connections is lost.

Via sizing and array placement, the distribution of decoupling capacitance across the tiers, backside power delivery, and electromigration budgeting are treated in detail in 3D Power Delivery.

Clock Distribution in 3D ICs

Vertical stacking offers a shorter global clock path and imposes a harder skew problem. A clock crossing between tiers passes through a via whose delay depends on geometry, on capacitance to the surrounding substrate, and on proximity to other vias, and small delay differences accumulate across several such transitions.

Two global strategies are common. The first replicates an independent clock network on each die and synchronizes at the die-to-die boundary, which decouples the tiers but requires clock-domain-crossing logic at every interface and pays a latency penalty there. The second distributes a reference through the stack and generates local clocks with a phase-locked or delay-locked loop on each die, which keeps the tiers in a common phase relationship but demands that the reference survive multiple via transitions with acceptable jitter. Wide parallel die-to-die interfaces sidestep much of this by forwarding a clock alongside the data it times, so that source and data share a common path and track each other's variation.

Temperature complicates skew control in a way it does not in a planar design. A stack has a vertical temperature gradient, so different tiers operate at different temperatures simultaneously, and both via characteristics and buffer delays shift with temperature. The resulting skew is not a fixed offset that can be trimmed once but a term that varies with workload. Adaptive distribution schemes use on-die temperature sensors and adjustable delay elements to compensate, holding skew within budget across the operating range.

Clock and power networks interact strongly. Clock buffers switch simultaneously and represent the largest coherent current transient on a die, and in a stack that transient must be supplied through the vertical power path. Supply droop translates directly into buffer delay and therefore into jitter, so the clock network's own current demand degrades the clock it distributes. Mitigation requires decoupling capacitance placed close to clock buffers, power via arrays positioned to serve them, and floor planning that keeps the impedance between the two low.

Jitter accumulates from intrinsic buffer noise, supply sensitivity, thermal noise, and coupling from neighboring signals, and the vertical path adds coupling opportunities that a planar route does not have. Dense via arrays place a clock transition physically alongside data transitions, so clock vias are commonly given dedicated positions ringed by ground vias for shielding. Differential distribution rejects common-mode disturbance on the most critical paths at the cost of doubling the vias those paths consume. The jitter budget must account for every contributor and still leave timing margin for the fastest interface in the system.

Thermal-Electrical Co-Design for 3D Systems

Stacking multiplies power density in a footprint while placing some dies farther from any heat sink. Heat generated in a buried die must conduct through the dies above it or through dedicated thermal paths, producing vertical gradients that raise junction temperatures well above what the same circuits would reach in a planar implementation.

The bonding interface is a significant part of the thermal resistance. A solder micro-bump joint surrounded by polymer underfill conducts heat poorly compared with the silicon on either side of it, so the interfaces themselves become the bottleneck in a tall stack. This is one of the underappreciated advantages of hybrid bonding: continuous metal and dielectric contact conducts far better than solder in underfill. Where micro-bumps remain, some architectures add thermal through-silicon vias that carry no signal and exist purely to conduct heat, sized and placed to balance thermal benefit against the die area and routing they consume.

Power delivery and heat removal share the same vertical structures, which couples two optimizations that would otherwise be independent. Resistive loss in the power path is itself a heat source, so undersizing the power vias raises both droop and temperature. Copper conducts heat as well as it conducts current, so power via arrays double as thermal conduits, and a design that sizes them for electrical margin often gains thermal margin at the same time.

Temperature dependence closes a feedback loop that makes analysis iterative. Conductor resistance rises with temperature, increasing droop and dissipation; transistor drive strength and threshold voltages shift, moving timing; and leakage current rises steeply with temperature, adding power that raises temperature further. Because each analysis changes the inputs of the other, thermal and electrical simulation must alternate until they converge, and the result must hold across the operating range rather than at one nominal point. Products complete the loop in hardware with distributed thermal sensors and control that throttles frequency, voltage, or activity when a sensor approaches its limit.

Known Good Die and Stack Yield

Bonding is permanent, so a defective die discovered after assembly destroys every good die bonded to it. This inverts the usual semiconductor test economics, in which comprehensive testing follows packaging, and it makes pre-bond screening—known good die testing—an architectural requirement rather than a manufacturing detail.

The arithmetic is unforgiving. If dies are screened to ninety-nine percent confidence and twelve of them are stacked, the stack yields roughly eighty-nine percent before any repair. Drop the per-die figure to ninety-five percent and the same twelve-high stack yields about fifty-four percent, with the cost of every scrapped stack including eleven good dies. Because yield loss compounds with height, the taller the target stack, the more test coverage each die must receive before it is committed.

Pre-bond testing faces a structural obstacle: the very structures that most need verification—through-silicon vias, bond pads, and die-to-die interface circuits—terminate on surfaces that become inaccessible once bonded, and their pads are too fine and too fragile for conventional probe cards. Practical approaches combine fine-pitch probe technology with built-in self-test structures that exercise vias and interface circuits from within the die, plus on-die measurement of via resistance and leakage. Post-bond testing must then verify the interconnect that pre-bond testing could not reach. IEEE Std 1838-2019, the standard for test access architecture in three-dimensional stacked integrated circuits, addresses this by defining die-level test features—a die wrapper register and primary and secondary test access ports—that compose into a stack-level architecture, so that each die and each inter-die interconnect layer can be tested individually both before and after stacking.

Redundancy recovers yield that screening alone cannot. Memory dies carry spare rows and columns substituted for defective elements, with the repair recorded in fuses or anti-fuses. Die-to-die interfaces carry spare lanes remapped during link training. Power networks carry redundant vias so that an open path degrades capacity rather than function. Repair must be coordinated across the stack, because remapping a lane on one die requires the die on the other side of the interface to agree.

Test thoroughness is ultimately an economic optimization rather than a technical maximum. Exhaustive characterization of every parameter on every die would cost more than the scrap it prevents, so test programs use sampling, statistical outlier detection, and adaptive coverage that responds to observed defect rates and failure analysis feedback. The right operating point depends on the relative cost of test time, of a scrapped die, and of a scrapped stack, and on the reliability the application demands—which is why an aerospace or automotive part and a consumer part reasonably choose different answers.

Heterogeneous Integration Challenges

The strongest argument for die stacking is that it lets each function use the process best suited to it. Logic benefits from the most advanced node available; dense memory, analog front ends, power management, and radio-frequency circuits often do not, and forcing them into an advanced node raises cost without improving them. Stacking lets a designer buy each capability where it is cheapest and combine them vertically. The difficulties are correspondingly varied.

Electrical compatibility comes first. A logic die in a leading-edge node may operate below one volt, while an adjacent analog, memory, or power management die requires substantially more. The interface must shift levels without adding meaningful delay or power, and it must tolerate the different drive strengths, parasitic loads, and noise sensitivities that different technologies present. Interface circuits also need protection appropriate to the assembly process, since electrostatic discharge requirements for a die that will be handled individually differ from those for a die bonded at wafer level.

Thermal mismatch is severe when power densities differ sharply. A processor die may dissipate tens or hundreds of watts next to a memory or sensor die consuming a small fraction of that. Vertical heat flow then pushes the low-power die outside the temperature range it was characterized for, which matters especially for analog and memory circuits whose behavior is temperature sensitive—dynamic memory retention time, for instance, falls as temperature rises, forcing more frequent refresh and consuming bandwidth. Placing the highest-power die nearest the heat sink is the obvious remedy, but that position may conflict with the electrically preferred arrangement.

Mechanical stress arises from mismatched coefficients of thermal expansion. Silicon dies bond to silicon dies without much difficulty because their expansion coefficients match. Combining silicon with compound semiconductors such as gallium arsenide or gallium nitride, or with materials chosen for thermal reasons, introduces differential expansion that loads the bonding interface through every thermal cycle and can cause fatigue cracking or delamination. Underfill selection, bonding temperature, and in some cases deliberately compliant structures manage this. Stress also has a purely electrical consequence: silicon is piezoresistive, so mechanical stress near a via shifts carrier mobility in nearby transistors, which is why keep-out zones around through-silicon vias exclude sensitive devices.

Manufacturing integration ties the rest together. Dies from different sources arrive with different test methodologies, different surface preparations, and different temperature tolerances, and the assembly flow must accommodate all of them. The bonding process cannot exceed the thermal budget of the most sensitive die in the stack. Supply chains must be coordinated so that dies with different lead times and yields arrive together, and the design should isolate the most expensive or least mature function so that a problem there does not scrap the rest. These are not signal integrity problems in themselves, but they constrain the architecture within which the signal integrity work is done.

Representative Implementations

High-bandwidth memory is the highest-volume application of true 3D stacking. A stack places several DRAM dies over a base die and links them with thousands of through-silicon vias, presenting an extremely wide, comparatively slow interface instead of a narrow fast one. The JEDEC HBM4 standard, JESD270-4, published in April 2025, defines a 2048-bit interface running at up to 8 Gb/s per pin for roughly 2 TB/s per stack, organized as thirty-two independent channels each carrying two pseudo-channels, and supports four-, eight-, twelve-, and sixteen-high configurations built from 24 Gb or 32 Gb dies. HBM4 retains solder micro-bumps rather than moving to hybrid bonding, a decision that illustrates how maturity and yield weigh against raw electrical advantage.

Processor cache stacking demonstrates the opposite choice. AMD's 3D V-Cache bonds an SRAM die directly onto the compute die using TSMC's stacked-die process at a bond pitch of roughly nine micrometers with no solder between the dies, adding cache capacity that would not fit in the compute die's own area while keeping the access path short enough that the added latency is small. Intel's Foveros family covers both approaches, with solder-based die stacking in shipping products and a bumpless Foveros Direct variant using direct copper bonding at sub-ten-micrometer pitch.

Stacked image sensors deserve mention because they were commercially significant before the current generation of logic stacking and because they illustrate wafer-to-wafer bonding at scale. Placing the pixel array on one die and the readout, conversion, and processing logic on another lets each layer use an optimized process, gives the pixel array the entire top surface, and shortens the path from pixel to converter. The dies are small and both layers are mature, which is exactly the condition under which wafer-to-wafer bonding's yield penalty is tolerable.

Conclusion

Die stacking has moved from a specialty technique to a mainstream architectural option, and the decisions it presents are genuinely architectural rather than merely mechanical. Whether tiers meet face to face or face to back, whether they are joined by solder or by direct copper contact, whether assembly happens at wafer or die granularity, and where the highest-power die sits in the stack all determine the electrical, thermal, and economic character of the finished system. None of these choices can be made in isolation, because the same vertical structures carry signals, supply current, and heat.

The direction of development is clear even where specific timelines are not. Interconnect pitch continues to fall as hybrid bonding matures, stacks continue to grow taller as dies are thinned further, and open die-to-die specifications continue to lower the barrier to assembling dies from different suppliers. As pitch approaches the scale of on-die wiring, the distinction between a die boundary and a routing layer erodes, and partitioning a design across tiers becomes a floor-planning decision rather than an interface design. Realizing that requires design methodologies that treat electrical, thermal, and mechanical behavior as one coupled system—which is the consistent lesson of every constraint described above.

Related Topics