日本語
한국어
CoWoS Packaging Guide for AI Systems and PCB Designers
KEY DEFINITION CoWoS Packaging is a 2.5D integration platform that places one or more logic dies and high-bandwidth memory on an interposer, then attaches the assembly to an organic package substrate. It shortens die-to-memory paths and raises bandwidth, while shifting demanding power, thermal, mechanical, and signal integrity requirements to the package and the system PCB.
Introduction
A modern AI accelerator is no longer a single chip sitting on a conventional package. The compute dies, stacked memory, interposer, package substrate, voltage regulators, and server PCB behave as one electrical and thermal system.
This guide explains how CoWoS Packaging works, how NVIDIA uses the approach, how the main variants differ, and what PCB engineers should verify before a high-power accelerator board reaches fabrication.
What Is CoWoS Packaging
CoWoS Packaging is a family of chip-on-wafer-on-substrate processes that integrates logic and high-bandwidth memory through a dense intermediate routing structure before the package is mounted on a PCB.
The word wafer refers to the stage at which known good logic and memory dies are joined to the interposer. The completed CoWoS Packaging assembly is then attached through larger C4 connections to an organic package substrate. That substrate provides the fanout to the ball grid array that connects the package to the system PCB.
In CoWoS Packaging, the 2.5D label means the logic dies and HBM stacks sit side by side on an interposer, even though each HBM stack is itself a three-dimensional structure connected with through-silicon vias. CoWoS Packaging therefore reduces the distance and energy required to move data between compute and memory without asking the PCB to route thousands of extremely fine die-level connections.
For a CoWoS Packaging board designer, the distinction matters. The interposer is not the PCB, and the organic package substrate is not the motherboard. The PCB still carries bulk power, management buses, high-speed fabric links, clocks, and mechanical loads. Its stackup and power distribution network must match the package vendor's land pattern, current profile, loss budget, and warpage limits.
How NVIDIA Uses CoWoS Packaging
NVIDIA uses CoWoS Packaging to place high-performance GPU logic next to HBM so the memory interface can be much wider and shorter than a conventional board-level interface.
The company publicly described CoWoS Packaging with the Tesla P100, where CoWoS and HBM2 delivered 720 GB per second of memory bandwidth. The integration model continued as accelerator memory expanded. NVIDIA specifies 80 GB of HBM3 and 3.35 TB per second for the H100 SXM, then 141 GB of HBM3e and 4.8 TB per second for the H200 SXM.
Blackwell-class modules raise the integration pressure again because multiple large logic dies, more HBM capacity, and a very fast die-to-die link must operate within one package. Current NVIDIA platform documentation lists 180 GB of HBM3e and up to 8 TB per second for the B200 SXM. CoWoS Packaging is valuable here because the interposer handles dense local connectivity while the PCB handles package power, external I/O, control, and cooling infrastructure.
CoWoS 2.5D Packaging Compared with Standard Packaging
CoWoS 2.5D Packaging differs from a standard single-die package mainly in integration density, memory proximity, routing scale, and the demands it places on the package substrate and host PCB.
| Design factor | Standard package | CoWoS 2.5D package | PCB implication |
|---|---|---|---|
| Die integration | Usually one main die | Logic dies and HBM share an interposer | Larger BGA, tighter escape planning |
| Local routing | Package substrate traces | Micron-scale interposer wiring | PCB routes external links, not HBM buses |
| Memory path | Longer and narrower | Short and very wide | More bandwidth reaches board interfaces |
| Power density | Moderate for many devices | Very high for AI accelerators | Low impedance planes and strong cooling |
| Mechanical risk | Smaller package area | Large multilayer stack | Warpage, coplanarity, and stiffener review |
| Test and cost | Simpler assembly flow | Known good die and compound yield dominate | Rework strategy must be agreed early |
CoWoS Packaging moves the finest routing inside the package, but it increases the importance of PCB power integrity, channel loss, assembly flatness, and thermal planning.
CoWoS S, CoWoS R, and CoWoS L Technologies
The three main CoWoS Packaging variants use different interposer constructions to balance routing density, package size, electrical performance, mechanical behavior, and manufacturing cost.
| Type | Interposer | Main strength | Published reference point | Best fit |
|---|---|---|---|---|
| CoWoS-S | Full silicon interposer with TSVs | Highest routing density and mature HBM integration | Up to 3.3 reticle size, about 2,700 mm2 | Dense high-end compute with a bounded interposer area |
| CoWoS-R | Copper and polymer RDL interposer | More compliant structure and larger scaling options | Minimum 4-micrometer pitch with 2-micrometer line and space | Large packages where RDL routing meets the electrical target |
| CoWoS-L | Molded RDL interposer with local silicon interconnects | Dense local links without a full-area silicon interposer | First 3.5 reticle implementation entered production in 2024 | Very large chiplet and HBM assemblies |
Within CoWoS Packaging, TSMC reports that CoWoS-R entered volume production in 2023 and that CoWoS-L reached volume production at 3.5 times reticle area in 2024. Those milestones show how CoWoS Packaging is expanding beyond the practical area limit of a full silicon interposer.
CoWoS Packaging Process
A representative CoWoS Packaging process joins tested dies to an interposer at wafer scale, completes the package on an organic substrate, and verifies the assembly before it reaches board-level manufacturing.
Known good die preparation: Logic dies and HBM stacks are electrically screened because one weak die can reduce the yield of the entire multichip package.
Interposer fabrication: The CoWoS Packaging process builds a silicon, RDL, or local silicon interconnect structure with the required routing, power distribution, and vertical connections.
Chip on wafer assembly: Logic and memory dies are aligned and attached through fine-pitch microbumps or related interconnects, followed by underfill or encapsulation as required.
Wafer-level inspection: Optical, electrical, and structural checks identify opens, shorts, alignment errors, voids, and other defects before expensive downstream steps.
Singulation and substrate attach: The integrated die and interposer unit is separated and connected to the organic package substrate through C4 bumps.
Thermal and mechanical completion: A thermal interface material, lid, stiffener, or heat spreader is added according to the module design and cooling method.
Final package test: Functional, burn-in, and reliability screens confirm the package before it is mounted to a carrier or server PCB.
PCB Inputs Required Before Layout
PCB layout for an accelerator that uses CoWoS Packaging should not begin from a BGA land pattern alone. The board team needs a controlled interface package with electrical, thermal, and mechanical limits.
Land pattern and keepouts: Obtain the final ball map, package outline, component height, underside restrictions, keepout zones, and allowable escape structures.
Power delivery data: Request rail currents, transient profiles, target impedance, permitted ripple, sense locations, and package-side decoupling assumptions.
Channel requirements: Lock differential impedance, insertion loss, return loss, crosstalk, reference plane, connector, and via budgets before choosing the stackup.
Mechanical limits: Review package warpage, PCB thickness, stiffener needs, heatsink load, assembly profile, and flatness requirements together.
Manufacturing controls: Define filled via-in-pad structures, microvia depth, backdrilling, copper balance, material availability, coupon design, and acceptance criteria with the fabricator.
Before releasing the PCB, ask PCBgogo for an early stackup, impedance, and DFM review so the proposed via structures and materials can be checked against an actual manufacturing process.
CoWoS Packaging Challenges
The main CoWoS Packaging challenges come from combining very large, valuable, and thermally intense components into one assembly whose yield depends on every interface.
Compound yield: A package may contain several logic dies and HBM stacks. Known good die screening, traceability, and test coverage are essential because a late defect can scrap a high-value assembly.
Warpage and CTE mismatch: Silicon, mold compounds, organic substrates, solder, and the PCB expand at different rates. Package size and thermal cycles can strain C4 joints, BGA joints, and the board unless materials and copper distribution are co-designed.
Power integrity: Fast accelerator load steps excite resonances across the voltage regulator, PCB planes, package substrate, and on-package capacitors. A useful PDN model must cover the complete path rather than treating the board and package separately.
Thermal density: High junction temperature reduces performance margin and reliability. Cold plates, heat spreaders, thermal interface materials, mounting pressure, airflow, and PCB temperature all need one verified thermal model.
Inspection and rework: Fine pitch internal joints are not accessible after assembly, and a large BGA has a narrow process window. X-ray inspection, coplanarity control, stencil design, reflow profiling, and an agreed rework policy become board-level necessities.
These risks explain why CoWoS Packaging programs require concurrent package, PCB, power, mechanical, thermal, test, and manufacturing reviews before layout is frozen.
Future Direction of CoWoS Packaging
The future of CoWoS Packaging is larger system integration with more HBM, more chiplets, denser local connections, and tighter electrical and thermal co-design from silicon through the PCB.
Larger interposers: TSMC states that 5.5 times the reticle area CoWoS Packaging was certified in 2025, with volume production planned for 2026. Larger footprints increase compute and memory capacity, but they also magnify warpage, yield, and substrate routing pressure.
HBM scaling: New HBM generations increase capacity and bandwidth while adding stack height, heat, and power delivery constraints. CoWoS Packaging must evolve with memory interfaces rather than treating HBM as a drop-in replacement.
Local silicon interconnects: Within CoWoS Packaging, CoWoS-L concentrates the finest wiring only where dies need it. This approach can support a larger molded interposer without paying for full-area silicon routing.
Three-dimensional integration: Future systems can combine 2.5D CoWoS Packaging with wafer-level logic stacking. That shortens critical links but makes thermal paths, test access, and known good die strategy more difficult.
Optical I/O: Co-packaged optics and silicon photonics can reduce the electrical reach between compute and the network. PCB designers will still manage power, control, fiber routing, connector access, and cooling around the optical engine.
Frequently Asked Questions
Is CoWoS Packaging a Type of PCB
No. CoWoS Packaging is an advanced semiconductor packaging platform that uses an interposer and an organic package substrate. The completed package is later soldered to a PCB, which provides system power, I/O, management, and mechanical support.
Is CoWoS Packaging the Same as Chiplet Design
No. A chiplet is a functional die used as part of a larger system, while CoWoS Packaging is one way to integrate chiplets and HBM. A CoWoS product may contain one large logic die or several logic chiplets.
Does Every NVIDIA GPU Use CoWoS Packaging
No. CoWoS Packaging is used where the performance and memory architecture justify advanced 2.5D integration, especially data center accelerators with HBM. Many graphics products use other package and memory arrangements.
Which CoWoS Packaging Variant Is Best
There is no universal best variant. CoWoS-S favors full silicon routing density, CoWoS-R uses a more compliant RDL interposer, and CoWoS-L combines an RDL structure with dense local silicon interconnects for large systems.
How Does CoWoS Packaging Affect PCB Stackup
It often pushes the PCB toward more deliberate power planes, low-loss signal layers, controlled return paths, advanced escape vias, and strict copper balance. The correct stackup comes from the package ball map, channel loss budget, current profile, board thickness, and fabricator capability.
Conclusion
CoWoS Packaging brings logic and HBM close enough to deliver exceptional bandwidth, but the result depends on coordinated control of interposer routing, package yield, PCB power integrity, channel loss, cooling, and mechanical reliability. Start the board stackup and DFM review while the package interface is still flexible, not after the ball map and enclosure are locked.
