# Selecting a bonding route under constraints

Canonical: https://semiaifoundry.com/learn/atlas/hbm-bonding-process-selection/
Revision: 2026-10-07
Rights: https://semiaifoundry.com/ai-atlas/licenses.html

Packaging · Atlas v1.3.0

# Selecting a bonding route under constraints

Compare qualified output, height, thermal behavior, inspection and cost on the same product boundary.

[Explore in Atlas](https://semiaifoundry.com/ai-atlas/#hbm-bonding-process-selection)Process-flow alternatives on a fixed stack/package boundary; schematic

Learning objectiveCompare qualified output, height, thermal behavior, inspection and cost on the same product boundary.

## The mechanism

A useful process comparison starts from one required product: die count, footprint, electrical interfaces, operating conditions and test coverage. TC-NCF, MR-MUF and hybrid bonding organize joint formation and support differently. Each route moves effort among surface preparation, die placement, joint formation, encapsulation, annealing and inspection. “Finer pitch” and “faster bond tool” are individual attributes, not complete cost or delivery answers. Keep development readiness and demonstrated results beside each route, and leave supplier yield, process time or commercial price unknown when undisclosed.

1.  Fix the same functional memory requirement and package-height boundary before comparing alternatives.
2.  Map the chosen route’s preparation, handling, bonding/fill, thermal and test steps without replacing them with one generic bond operation.
3.  Define compatible station capacity and conditional yield for the same input population.
4.  Include consumables, equipment occupancy, inspection and losses in a cost per accepted stack model.
5.  Reject or revise routes that miss electrical, thermal, mechanical or qualification requirements even if nominal throughput is high.

## Science and assumptions

### Cost per accepted stack

$$
c_{\mathrm{good}}=\frac{c_{\mathrm{input}}+c_{\mathrm{process}}+c_{\mathrm{test}}}{Y}
$$

cgood = (cinput + cprocess + ctest) / Y

Symbols and units

cgood

Average modeled cost per accepted stack

currency/stack

cinput

Screened components cost per starting set

currency/set

cprocess

Assembly cost per starting set

currency/set

ctest

Inspection/test cost per starting set

currency/set

Y

Accepted stacks per starting set

fraction

Costs are normalized to one assembled input set and divided by final pass fraction.

Assumptions Simple no-rework average with all modeled costs incurred per input set; stage-dependent scrap/rework requires a yield waterfall.

## Engineering example

Synthetic routes: 100 currency units per input set at 90% final yield cost 111.1 per accepted stack. A 105-unit route at 97% costs 108.2. The example illustrates sensitivity; it ranks no commercial process.

## Distinctions and tradeoffs

A process must close the full product and production requirements; one favorable parameter does not establish a usable stack.

Reducing joint thickness can consume more surface-control effort; reducing bond time may expose inspection or supply as the next bottleneck.

Memory economics and supply depend on accepted output from the complete route, affecting how many accelerators can be deployed.

## Specifications and interfaces

Comparison basis

Same product and test coverage

Different layer counts or qualification stages are not an equal comparison.

Readiness

Research, sample, qualified, production

Keep date and measured scope visible.

### Interfaces

Process candidate Input

Flow, equipment and material requirements

Integration decision Output

Feasible route plus remaining evidence needs

## Check your understanding

### Which comparison supports an engineering choice?

1.  Same product boundary, yield denominator and qualification requirements
2.  Only minimum reported pitch
3.  Only the fastest advertised bonding operation

Answer and reasoning

Same product boundary, yield denominator and qualification requirements

A process must close the full product and production requirements; one favorable parameter does not establish a usable stack.

## Evidence and scope

Route mechanisms are grounded in their named sources. Cost, cycle time and yield values in the workbench are transparent teaching scenarios, not supplier disclosures.

Adjust stack layers, thickness, joint route, interface bandwidth and conditional yield; presets are synthetic, not supplier process windows.

## Sources

-   [Samsung — 36 GB HBM3E 12H development](https://news.samsung.com/global/samsung-develops-industry-first-36gb-hbm3e-12h-dram)
-   [SK hynix — Small Size, Big Impact](https://news.skhynix.com/en/small-size-big-impact/)
-   [imec — wafer-to-wafer hybrid bonding at 400 nm pitch](https://www.imec-int.com/en/articles/wafer-wafer-hybrid-bonding-pushing-boundaries-400nm-interconnect-pitch)

[Published lesson data](https://semiaifoundry.com/ai-atlas/lesson-data/hbm-bonding-process-selection-35150926514801ce.json) · [Atlas releases and licensing](https://semiaifoundry.com/ai-atlas/licenses.html)

## Continue exploring

### Foundations

-   [Thermocompression with non-conductive film](https://semiaifoundry.com/learn/atlas/hbm-tc-ncf-route/)
-   [Mass reflow and molded underfill](https://semiaifoundry.com/learn/atlas/hbm-mr-muf-route/)
-   [Hybrid bonding: surface-to-interface closure](https://semiaifoundry.com/learn/atlas/hbm-hybrid-bond-route/)

### Related lessons

-   [HBM silicon intensity and allocation](https://semiaifoundry.com/learn/atlas/hbm-silicon-intensity/)
-   [Compatible supply and converging bottlenecks](https://semiaifoundry.com/learn/atlas/compatible-supply-capacity/)
-   [Package and system preserve the contract](https://semiaifoundry.com/learn/atlas/crg-package-system-closure/)