Skip to main content

Overview

General comment

HBM4 16-high 64 GB stack – max-density SKU.

Closes a JEDEC-generation coverage gap.

All three HBM4 stack heights are built on one declared core die (11 x 11 mm, 121 mm², 1γ) carrying 32 Gbit, so per-die capacity is identical across the axis and only the die count changes: 32 Gbit on 121 mm² is 0.264 Gb/mm², a 10% gain on the HBM3E core die (24 Gbit on 99.75 mm²) and below the 0.412 Gb/mm² 1c array density an HBM core die cannot reach because of its TSV farm and per-channel logic. System boundary: this dataset ends at the assembled and tested stack. Silicon interposer, 2.5D redistribution (CoWoS-class), organic package substrate, and package-level assembly are deliberately excluded - those are owned by the consuming 2.5D-packaging dataset under the count-once convention, so composing this stack with a package dataset does not double-count. Accelerator/module-level assembly is likewise outside this dataset's boundary and is not modeled in REEL v1.0. This variant represents early-ramp HBM4 production: HBM4-bound 1c yields reported <60% as of Apr 2026; Samsung 1c ~65-70% general.

Technology
HBM4 16-high, 64 GB
Geography
Global average – Samsung (Korea), SK Hynix (Korea), Micron (US/Japan/Taiwan); base die from TSMC (Taiwan)

System boundary

System boundary - HBM4 64GB, 16-Hi Stack, ramp yield scenarioSystem boundary figure: identity, gate in, entering flows, the dashed system boundary and the unit processes inside it, the reference product, emissions and waste, and below it the flows that are recorded but not quantified. No inventory quantities.HBM4 64GB, 16-Hi Stack, ramp yield scenarioGATE INIncoming polished DRAM and base-die wafers(wafer fab included)ENTERING FLOWSMaterialsFreight transportOtherSYSTEM BOUNDARYCOMPOSED OF OTHER REEL DATASETSDRAM Core Dies1-gamma DRAM wafer, 300mm, moderatescenarioBase Logic Die5nm FinFET wafer, 300mm, Taiwan,moderate scenarioTSV StackingHBM4 TSV Stacking and Assemblypackage, 16-HiREFERENCE PRODUCTHBM4 64GB, 16-Hi Stack, ramp yieldscenarioAssembled and tested memory component(excludes interposer, package substrate,and module-level assembly)EMISSIONS AND WASTEEmissions to airEmissions to waterWaste routesRECORDED BUT NOT QUANTIFIEDIn scope, left out of the quantified inventoryCut off below the significance threshold, or with no background dataset available - each is listed with its reason1FLOWsystem boundaryreference flowentering flowemission / waste

The figure groups this dataset's unit processes by class. It is not a count of manufacturing steps — each process runs over as many passes as the flow requires, and those pass counts ship with the dataset.

Download this figure (SVG)

Data quality and references

Composite DQI 2.3 Good
Reliability
3.0
Completeness
2.1
Temporal
2.0
Geographic
2.0
Technological
2.4
How the score is calculated: Constituent datasets have more influence when they contribute a larger share of the modelled cradle-to-gate energy demand.

Pedigree scores follow the ecoinvent data-quality matrix: 1 is the best attainable, 5 the weakest. The composite is their aggregate.

Sampling procedure
JEDEC JESD270-4A public summary, vendor press releases (Samsung, SK Hynix, Micron HBM4 16-Hi pilot), ISSCC 2026 Samsung HBM4 paper (via SemiAnalysis), Wing Venture Capital deep dive, Korea Herald 16-layer disclosure, TrendForce 16-layer roadmap and Samsung hybrid pilot, Tom's Hardware on Samsung hybrid bonding, NomadSemi MR-MUF/TC-NCF yield analysis, and the REEL DRAM core-die, logic base-die and TSV-stacking datasets this dataset links as inputs.
Coverage status
Composed from linked REEL datasets
Pedigree-scored source files
49 — the source records behind this dataset's manufacturing operations. Each carries the five pedigree axes above; the composite DQI aggregates them.

Technosphere inputs

7 flows. Quantities are not published; they ship with the dataset on Circa.

DRAM Core Dies REEL datasetwafer
Derivation basis
  • Calculated by the manufacturing model. The linked REEL dataset carries the upstream life cycle sources.

No source is attached to this row.

Source citations
Not stated
Background dataset
Modelled by REEL; see the upstream dataset above.
Notes
Wafer demand calculated from die count, rectangular wafer packing, die yield and assembly yield. Wafer manufacturing yield is already included upstream.
Uncertainty
No range defined.
Unit
wafer
Base Logic Die REEL datasetwafer
Derivation basis
  • Calculated by the manufacturing model. The linked REEL dataset carries the upstream life cycle sources.

No source is attached to this row.

Source citations
Not stated
Background dataset
Modelled by REEL; see the upstream dataset above.
Notes
Wafer demand calculated from die count, rectangular wafer packing, die yield and assembly yield. Wafer manufacturing yield is already included upstream. The sort yield that sizes this row is not published for this die on its own, so it follows the sort yield of the memory core dies stacked above it; a scenario that varies the core-die yield therefore moves this row with it.
Uncertainty
No range defined.
Unit
wafer
TSV Stacking REEL datasetpackage
Derivation basis
  • Calculated by the manufacturing model. The linked REEL dataset carries the upstream life cycle sources.

No source is attached to this row.

Source citations
Not stated
Background dataset
Modelled by REEL; see the upstream dataset above.
Notes
Adjusted for packaging yield.
Uncertainty
No range defined.
Unit
package
Cu pillar + SnAg solder cap Materialkg · -55.6% / +77.8%
Derivation basis
  • Calculated from per-operation consumption, operation counts, and manufacturing yield.

No source is attached to this row.

Source citations
Not stated
Background data
composite
Background dataset
Modelled as a composite of its constituent materials.
Uncertainty
-55.6% / +77.8% around the published quantity. The bounds themselves ship with the dataset on Circa.
Unit
kg
Molding Compound Materialkg
Derivation basis
  • Calculated from per-operation consumption, operation counts, and manufacturing yield.

No source is attached to this row.

Source citations
Not stated
Background data
ecoinvent 3.12
Background dataset
epoxy resin insulator, SiO2
Uncertainty
No range defined.
Unit
kg
Wafer transport, KR fab to packaging site - Road Freight transporttkm · -100% / +172%
Derivation basis
  • Specified for this manufacturing operation and scaled to the dataset's functional unit.

The sources below come from the manufacturing operations behind this row.

Sources (inherited)
Inherited, rolled up from the contributing process steps:
  • Additional Investment in Onyang Campus... Expanding Back-End Processing Too (Seoul Economic Daily)
  • The Elec report, 2026-07: Samsung back-end investment focused on high-bandwidth memory
  • DigiTimes report, 2024-11-13: Samsung Cheonan and Onyang back-end hubs
Background data
ecoinvent 3.12
Background dataset
transport, freight, lorry, unspecified
Notes
Freight for the part-finished product moving between REEL-modelled sites: the wafer fraction this dataset consumes, travelling from memory fab to domestic back end. Shipped mass is that wafer fraction times the wafer mass implied by its declared diameter, times a shipping-packaging factor covering the returnable front-opening shipping box and its secondary carton; the packaging production burden is excluded as returnable, while its mass is carried because freight scales with what is actually shipped. Road freight is allocated on actual shipped mass, and the mapped road market already includes carrier empty returns; the central therefore has no return uplift. The authored band uses the route-class distance envelope, and its upper endpoint also adds the separate reusable-carrier return service.
Uncertainty
-100% / +172% around the published quantity. The bounds themselves ship with the dataset on Circa.
Unit
tkm — tonne-kilometres, a transport effort rather than a mass
Wafer transport, TW fab to packaging site - Air (short haul) Freight transporttkm · -47.4% / +80%
Derivation basis
  • Specified for this manufacturing operation and scaled to the dataset's functional unit.

No source is attached to this row.

Source citations
Not stated
Background data
ecoinvent 3.12
Background dataset
transport, freight, aircraft, dedicated freight, short haul
Notes
Freight for the part-finished product moving between REEL-modelled sites: the wafer fraction this dataset consumes, travelling from foundry logic die to memory stacking site, cross-border. Shipped mass is that wafer fraction times the wafer mass implied by its declared diameter, times a shipping-packaging factor covering the returnable front-opening shipping box and its secondary carton; the packaging production burden is excluded as returnable, while its mass is carried because freight scales with what is actually shipped. Air freight is allocated on chargeable weight rather than actual shipped mass. The authored band holds the fixed route distance: its lower endpoint uses actual shipped mass, its central uses chargeable-weight allocation, and its upper endpoint adds the separate reusable-carrier return service, not aircraft backhaul.
Uncertainty
-47.4% / +80% around the published quantity. The bounds themselves ship with the dataset on Circa.
Unit
tkm — tonne-kilometres, a transport effort rather than a mass

Outputs and waste

Failed packages (assembly reject) Solid wasteg · -3.8% / +7.9%
Derivation basis
  • Calculated from the mass balance of spent materials and consumables, with treatment selected from the waste classification.

No source is attached to this row.

Source citations
Not stated
Background data
ecoinvent 3.12 treatment route
Background dataset
treatment of waste electric and electronic equipment, shredding
Uncertainty
-3.8% / +7.9% around the published quantity. The bounds themselves ship with the dataset on Circa.
Unit
g

Emissions to air

No emissions to air are recorded at this level.

Emissions to water

No emissions to water are recorded at this level.

Flows not quantified

1 flow is recorded for this dataset but carries no quantity — cut off below the significance threshold, or with no background dataset available. It remains inside the declared system boundary; each is listed with its reason.

Not quantified - no background dataset available (1)

NCF Adhesive Materialkg · ±4.3%
Derivation basis
  • Calculated from per-operation consumption, operation counts, and manufacturing yield.

No source is attached to this row.

Source citations
Not stated
Background data
no background dataset
Background dataset
Not applicable: this flow is not quantified in the inventory.
Reason
A specialty thermosetting epoxy bonding film used in wafer-level stacking. No production dataset for it exists in the background database used for linkage; plain epoxy resin is the closest constituent but leaves out the cure system that defines this film, so it is not substituted. The quantity is recorded here and its upstream production is left unlinked.
Uncertainty
±4.3% around the published quantity. The bounds themselves ship with the dataset on Circa.
Unit
kg

Limitations and unquantified flows (6)

Limits this dataset declares about itself: first any limit stated in its own description, then any limit it declares flow by flow, grouped by channel. Each entry below is the model's own disclosure.

From the dataset description

  • PROSPECTIVE DATASET – 16-high HBM4 has not yet entered mass production, and no production data for it is public.
  • Wide uncertainty (+40% / -25%) reflecting first-generation yield risk, estimated die area, and an under-counted packaging bill of materials.

General

  • The mould-compound mass booked on this stack is an authored estimate that the stack's own packaging inventory contradicts. Three separate figures for the same encapsulant are in play across the files that describe this stack: the mass booked here as purchased demand, the mass the stacking inventory derives from the package envelope and the die stack inside it when it prices a scrapped stack, and the encapsulant basis the mould-waste allocation uses to size the flash it trims. They do not agree, and the flash waste this dataset ships is sized on the third of them. Reconciling them needs a package height and a base-die thickness that no file on this path declares consistently across stack heights - above the reference stack height the declared die and gap thicknesses no longer fit inside the declared envelope - and no public dimension for the taller rungs was found. The mass is therefore left as authored and disclosed here rather than re-derived on an assumed envelope.
  • The post-mould grinding slurry leaves this stack as waste with no matching input. It is a consumable of the moulding cycle, and the inventory that books its disposal does not book its purchase, so the mass it carries out is not balanced by a mass carried in. The single-use release film beside it, which had the same gap, is now booked as an input at the mass it leaves as waste. The slurry is not, because that same inventory records an unresolved question about whether the raw slurry and the treatment residues already booked as outputs in this dataset describe the same material twice; booking the raw slurry as an input before that is settled would compound it rather than close it.
  • The microbump mass on this stack is proportional to the square of the bump pitch this dataset declares, and that pitch is not traced to a vendor statement - it is an authored design point. The derivation is sound on the declaration, and the declaration is the weaker half: a coarser pitch at the same bump count would raise the metal mass in proportion. The pillar diameter behind the derivation is not declared on this dataset either; it is taken as the same fraction of the pitch that the high-bandwidth flash stack in this family declares, and the band on the row spans a pillar from half the pitch to one touching its neighbour.
  • The microbump count on this dataset is derived, not declared by a vendor: this generation's files state no count, only that it follows the doubled interface width, so the count is carried over from the previous generation on this path at the same number of bumps per data bit. The bump metal mass is proportional to it.

The flows above are this dataset's own records. What follows are the scope rules set once for the whole database in the methodology report, repeated here so every dataset page carries them.

Database-wide boundary policy — applies to every REEL dataset

These boundaries are set once for the whole database, in Chapter 2 of the methodology report, and apply to this dataset wherever they are relevant to it. The excluded flows listed above are specific to this dataset.

Use phase
Product operation is outside the cradle-to-gate scope.
End-of-life treatment
Recycling and disposal are outside the cradle-to-gate scope.
Distribution and retail
The gate is a finished component ready for integration into a higher-level assembly.
Inbound transport of raw materials
Transport of purchased raw materials to the manufacturing facility is already inside the upstream "market for" datasets that users link to a background database, so it is not modelled a second time here. This does not cover freight between REEL production stages, which is modelled where a dataset authors it.
Returnable shipping containers
Where freight between production stages is modelled, the mass moved is the product itself. The shipping container (FOSB, SEMI M31) is returnable capital equipment whose per-trip share is unsourced, so its tare is excluded from the transport effort.
Photomask fabrication
A mask set's embodied burden is amortised across a high-volume production run and is not attributed per wafer. Users assessing low-volume production should add mask fabrication separately; the methodology report gives the basis for the exclusion.
Employee transport, administration and R&D overhead
Employee transportation, facility administration and R&D/pilot-production overhead are outside scope.
Capital goods
Manufacturing equipment, cleanroom construction and facility infrastructure are excluded, on the grounds of absent public data on equipment embodied energy, uncertainty in equipment lifetime and allocation, common practice in electronics LCA, and a focus on the operational inventory. Future versions may include capital goods when sufficient public data becomes available.
Precious metal recovery credits
Scrap recovery credits for precious metals are excluded pending data availability.
Wafer reclaim
Test wafers and scrap are outside the system boundary.

Inside the boundary, linked rather than modelled

Silicon ingot growth and wafer slicing
Inside the cradle-to-gate scope, but treated as upstream material inputs linked to background databases rather than modelled as REEL processes.
Freight between production stages
Where a dataset's product moves between REEL production stages - wafer fabrication to the packaging site, for example - that leg is authored as a transport service and linked to an ecoinvent freight activity. It is measured as a transport effort in tonne-kilometres, not as a mass. Route distances are authored per route class with a stated band; a lower bound of zero is a modelling statement that the two sites can be co-located, not a missing value.

Cut-off criteria

A flow is excluded from a process inventory when it contributes less than 1 % of the total mass of inputs to that unit process, or less than 1 % of its total energy input. The denominator is total process inputs, not product mass. That distinction matters in semiconductor manufacturing, where the input mass of water, chemicals and gases greatly exceeds the product mass, so the threshold removes only genuinely minor flows.

Included regardless of the cut-off

  • Perfluorocarbons (CF4, C2F6, SF6, NF3) - high GWP, EPA regulated
  • Heavy metals (Pb, Cd, Hg, Cr(VI)) - RoHS regulated, high toxicity
  • Volatile organics (photoresist solvents, PGMEA) - air quality
  • Precious metals (Au, Ag, Pd, Pt) - high embodied impacts
  • Ozone-depleting substances (legacy CFCs, HCFCs) - Montreal Protocol