The mechanism
Every added die introduces heat sources, interfaces and opportunities for assembly damage. Heat follows several parallel paths through silicon, metal, polymer and the surrounding package; the hottest location is not necessarily the tallest die. Test coverage changes with access: a wafer test can inspect some die functions before stacking, while TSV chains, cross-die connections and the complete high-speed interface require later checks. A known-good input is a unit that passed a defined screen, not a physically perfect die. Assembly yield therefore starts from screened inputs and counts newly introduced failures, avoiding a second charge for the same wafer loss.
- Define incoming wafer/die test coverage and preserve identity through thinning, singulation and stacking.
- Measure continuity and functional behavior at the stages where test access exists; map errors to the same stack genealogy.
- Estimate heat from each active layer using a stated power distribution and boundary temperature.
- Inspect assembly defects with electrical tests plus suitable imaging; a hot spot or open circuit alone does not identify its cause.
- Separate per-unit screening, product reliability qualification and observed in-service failure rates.
Science and assumptions
Multiple heat sources
Ti − Tcoolant = Σj Rij Pj
Symbols and units
- Ti
- Temperature at inspected location i
°C
- Tcoolant
- Specified coolant boundary temperature
°C
- Rij
- Thermal coupling from source j to location i
K/W
- Pj
- Heat dissipated at source j
W
A thermal-resistance matrix represents how each heat source j contributes to temperature at location i, including cross-heating.
Assumptions Linear steady-state approximation at specified boundaries. Transients and temperature-dependent properties require a richer model.
Conditional assembly yield
Yassembly ≈ yjoint^J
Symbols and units
- Yassembly
- Probability a screened input set survives modeled assembly
fraction
- yjoint
- Conditional aggregate pass probability per interface
fraction
- J
- Modeled interfaces
count
J is the number of independently modeled assembly interfaces, each with aggregate conditional pass probability yjoint.
Assumptions Screened inputs; excludes residual input escapes, correlated defects, repair and common tooling failures. An interface probability is not a per-bump probability.
Engineering example
Synthetic independent-interface model: 12 interfaces at 99.5% conditional pass probability yield about 94.2%; at 99.9%, about 98.8%. A shared contamination event violates independence. These values do not estimate any supplier’s yield.
Distinctions and tradeoffs
Track the denominator at each stage; conditional assembly loss acts on the units that actually enter assembly.
Earlier screening avoids wasting downstream components, but adds handling, test time and cost; some failures are only observable after assembly.
Effective stack supply and sustainable frequency depend on both manufacturability and thermal operation.
Specifications and interfaces
- Yield denominator
- Screened input sets
Do not multiply their original wafer yield a second time.
- Thermal boundary
- Location, power map and coolant
Stack-average temperature can hide local limits.
Interfaces
- Die screening Input
- Coverage, bin and residual escape risk
- Stack test Output
- Functional, interconnect and thermal evidence
Check your understanding
Why not multiply wafer yield into every already screened die again?
- It double-counts losses already excluded from the incoming population
- Screening guarantees zero future failures
- HBM assembly cannot introduce faults
Answer and reasoning
It double-counts losses already excluded from the incoming population
Track the denominator at each stage; conditional assembly loss acts on the units that actually enter assembly.
Evidence and scope
imec’s 2025 thermal study concerns a specified research HBM-on-GPU arrangement. It motivates spatial heat analysis; its simulated temperatures are not general product limits.
Adjust stack layers, thickness, joint route, interface bandwidth and conditional yield; presets are synthetic, not supplier process windows.