Assumptions and defaults register

This page centralizes shared numerical assumptions and conventions. It is the single reference register for defaults. A default repeated across chapters must appear here.¹

Usage rules

• This register records assumptions shared by multiple chapters.
• Local pages may repeat a value without creating a competing register.
• Associate entries with a source, date and application context.
• Justify revisions; see Time and versioning.

Shared assumptions

General framework

Assumption	Value/convention	Source	Application
Methodological perspective	Attributional by default	Digital4Better methodology	All assessments; consequential variants discussed where relevant
Environmental indicators	Nine midpoints plus electricity consumption	D4B convention; see Indicators	All assessments, subject to available compatible factors
Source hierarchy	Measurement → primary → provider → literature → generic estimate	Source hierarchy	Selection of assessment data

Electricity and energy

Assumption	Value/convention	Source	Application
Electricity factors	digital4better/data preferred	data/factor/*, data/energy/*	Most robust compatible time resolution; monthly where available, annual otherwise 2
Electricity mix	Consumption-based, including imports/exports	digital4better/data	Operational impacts; see the cloud convention for the renewable share
Generation technologies	Coal, gas, other fossil, wind, solar, bioenergy, hydro, other renewable, nuclear	UNECE; Energy Conversion and Management, 2022	Country-factor calculation

Datacenters and infrastructure

Assumption	Value/convention	Source	Application
Default PUE	1.58, historical scenario	Historical Uptime Institute reference; not a verified current mean	When actual PUE is unknown 3
RAM power	0.000392 kW/GB	Cloud Jewels / Cloud Carbon Footprint	Server component estimate 4
SSD power	1.2 × 10⁻³ kW/TB	Cloud Jewels / Cloud Carbon Footprint	Server component estimate
HDD power	0.65 × 10⁻³ kW/TB	Cloud Jewels / Cloud Carbon Footprint	Server component estimate
Inter-datacenter network	Scenario: 0.001 kWh/GB	Cloud Carbon Footprint Appendix I; inconsistent Gb/GB notation in prose requires clarification	Does not automatically cover LAN or user access
Server lifetime for reference VM factors	Five years	NegaOctet / Base Empreinte	Embodied VM factors

Networks

Assumption	Value/convention	Source	Application
D4B network factors	Per-GB·km factors, or per-byte·km before conversion	D4B convention still in use; network model	Volume × estimated distance × factor; per-GB illustrations incorporate the route upstream
Within-country distance	Estimated country radius: $r=\sqrt{A/\pi}$	Country-area data and D4B convention	Approximation of national segments
Between-country distance	Great-circle route between centroids	Centroid data and spherical geometry	Approximation of international segments
Home-router traffic	220 GB/month	Historical ICT 2020 and ADEME/ARCEP 2022 context	Router allocation within fixed-network factors
CDN cache hit rate	70–90% sensitivity scenarios; historical web example 95%	Unmeasured D4B assumptions	Reduction of traffic after the CDN, not the whole user path

Component embodied models

Component	Equation	GWP coefficients	Source/status
Historical CPU, not complete GPU card	$I=(die_{mm^2}C_{die}+C_{base})F_{node}F_{pkg}$	$C_{die}=0.0197$ kg CO₂e/mm²; $C_{base}=9.14$ kg CO₂e	Historical UBA/Boavizta coefficients; node/package factors are D4B scenarios
RAM	$I=((capacity_{GB}/density)C_{die}+C_{base})F_{pkg}$ per module	$C_{die}=2.20$ kg CO₂e/cm²; $C_{base}=5.22$ kg CO₂e; scenario density 0.625 GB/cm²	UBA historical model; density is not universal
SSD	$I=(capacity_{GB}/density_{NAND})C_{die}+C_{base}$ per drive	$C_{die}=2.20$ kg CO₂e/cm²; $C_{base}=6.34$ kg CO₂e; scenario density 48.5 GB/cm²	UBA historical model; density is technology-dependent
HDD	$I=31.11$ kg CO₂e per drive	Generic scenario	Green Cloud Computing, UBA, 2021
Display	$I=area_{cm^2}C_{technology}+C_{base}$	LCD 0.07 kg CO₂e/cm²; OLED 0.12; base 3.0 kg	D4B educational assumption; calibration not established
Battery	$I=capacity_{kWh}F_{chemistry}$	LCO 75; NMC 65; LFP 50 kg CO₂e/kWh	D4B scenarios, not validated by the cited battery publications

Semiconductor multicriteria reconstruction

Status: not validated on 16 October 2026. Previous AP, WU, PM, IR and ADPe values reconstructed from an Intel inventory are no longer defaults. Reintroduction requires elementary flows, units and compartments, versioned LCIA factors, surface/yield rules and a reproducible calculation. Replaced values remain in refactoring archives.

CPU, RAM and SSD models retain explicitly sourced historical coefficients. Physical water cannot be added to characterized water impacts in m³ equivalents.

Manufacturing-country scenarios

These countries are sensitivity scenarios, not known locations for every component. Aggregate manufacturing factors cannot be relocated by multiplying by an electricity-mix ratio. Substitution requires an inventory separating electricity from other processes.

Component	Scenario country	Status
Advanced logic	Taiwan	D4B assumption; actual fabrication site remains to be established
RAM and NAND	South Korea	D4B assumption; actual fabrication site remains to be established
Assembly, displays, batteries, PCBs	China	D4B assumption; actual manufacturing site remains to be established

CPU/GPU process-node scenarios, $F_{node}$

These are sensitivity assumptions without an identified reproducible imec calibration. They do not demonstrate the environmental performance of a chip generation.

Node	$F_{node}$	Status
28 nm	1.0	D4B reference scenario
14 nm	1.2	Exploratory D4B assumption
10 nm	1.4	Exploratory D4B assumption
7 nm	1.8	Exploratory D4B assumption
5 nm	2.2	Exploratory D4B assumption
4 nm	2.5	Exploratory D4B assumption
3 nm	2.8	Exploratory D4B assumption

Equipment and lifetime

Assumption	Value/convention	Source	Application
Preferred approach	Component reconstruction where justified by data; documented equipment factors as an alternative	Digital4Better methodology	Representativeness takes precedence over apparent detail
Missing data	Measurement → primary → provider → literature → generic estimate	Source hierarchy	Missing parameters
Lifetime adjustment	$IF_{emb}=TIF_{emb}\times TL/RL$	D4B convention	Adjustment of a time-normalized reference factor when actual and reference lifetimes differ
vCPU	Relationship to hardware threads depends on offering and architecture	D4B allocation convention	Explicit mapping required for component allocation

Current exclusions

Excluded topic	Reason	Future work
Videoconferencing	Outside this iteration	Possible later use case
Video streaming	Outside this iteration	Possible later use case
IoT / embedded systems	Outside this iteration	Possible later use case

Historical web-service scenarios

The 69/34 seconds per view and 59% mobile values come from historical references in the web chapter (2023). The historical score uses a different convention: 20 seconds and 58% mobile. None is a guaranteed current average. The internal view-distribution model and conversion of one user-second to one vCPU-second are not validated out of sample and do not replace service measurements. December 2023 tables remain for traceability, with method and unit qualifications.

Simulators and validation snapshot ; 16 October 2026

Parameter	Value/rule	Status and provenance
Simulator electricity	2025 annual averages, 15 countries in local file	D4B commit 77cae4a5, data/factor/country-yearly.csv; local selection in src/data/electricity-factors.json
REF	No automatic reduction of reference mix	Separate sensitivity only; SCI compatibility requires verification
WUE	L/kWh IT × IT energy	Physical direct water; distinct from characterized WU
IR	Numerical total withheld from equipment simulator	Resolve historical kg/kBq inconsistency from upstream inventories
16:9 display	Area = diagonal² × 2.54² × 144/337	Exact geometry; diagonal in inches, area in cm²
AI compute	2P per input/output token; 6P per training token, active P	Dense-layer approximation; attention, memory and communication require calibration
AI fine-tuning	Training-scale compute scenario	No saving inferred from LoRA trainable fraction alone
AI MFU / initial PUE	0.4 / 1.2	Scenario assumptions, not provider measurements
AI lifetime	5 × 365.25 × 24 hours	Simulator allocation by active time; reserved-time allocation remains separate
AI cache	Residual cost α = 0.1; reuse 0–100%	Assumption applied before time/energy; memory overhead not measured
AI hardware	Component power and manufacturing in src/data/ai-hardware-profiles.json	Scenarios, not certified LCAs; B300 host memory corrected to 2 TB
Battery	Form: LCO 75, NMC 65, LFP 50; documented NMC 811 scenario: 60 kg CO₂e/kWh	Scenarios; water/ADPe withheld without traceable factors

Local capacity, lifetime, load, density and composition presets are visible in each form. They are editable examples, not equipment recommendations. Process and packaging coefficients remain exploratory until calibration is published.

Dataset gaps

Further evidence is needed for:

• explicit critical-material inventories.
• upstream semiconductor modeling conventions.
• consistent multicriteria factors for some supporting layers.
• embodied LAN factors in the system approach.

Assumptions and limitations

• The register is being consolidated and must be updated when a shared assumption changes.
• Defaults are working conventions, not universal facts. Better-matched evidence replaces them when available.
• An unavailable factor is not zero. An exploratory coefficient is not a measured value.

Background for data selection, not validation of default values.

Component illustration parameters

The shared register is versioned d4b-components-illustration-2026.10.19. It separates historical coefficients from the assumptions below. Centralisation makes examples consistent; it does not establish scientific calibration.

Parameter	Exercise value	Scope
DDR5 RAM	Density 1 GB/cm²; packaging 1.1	Same formula for RAM, terminal and server examples
HBM2e / HBM3	Densities 0.5 / 0.8; packaging 3 / 3.5	Exploratory scenarios; interposer counted once
Server memory modules	Assumed 32 GB per module	Count rounded up; total capacity preserved
Server drives	Assumed 2 TB SSD, 4 TB HDD	Same rule; per-drive constants follow drive count
Structural residual	Smartphone 4, tablet 5, laptop 8, desktop 12, server 50 kg CO₂e	Assumptions without overlap with inventoried components
Device power	50% CPU TDP + display + 2 W auxiliary	Exercise model, not measured fleet-average power
Server power	CPU/GPU TDP + RAM + storage + 50 W auxiliary	Constant-power scenario before PUE
Computer mix	60% laptop, 40% desktop with display	Active-time weighting, not audience statistics
Mobile mix	95% smartphone, 5% tablet	Same convention
Device profile allocation	Calculated manufacture / lifetime / 8,760 h/year	Per-second factor is derived, not copied from old flat factors

Configurations also describe capacity, component count, display and battery. Unknown properties of actual equipment remain assumptions or unassessed items. Old VM and device factors are historical references, no longer the target model.

Service device profiles use 365 days per year for the per-second factor. Component examples and AI use 365.25 days for annual energy or time allocation. These conventions remain explicit; harmonising them would change results and require a new version.

Footnotes

1. digital4better/data. Methodological corpus. ADEME Base Empreinte. Digital4Better data catalogue, snapshot 77cae4a. References and scope: Sources and corpus. ↩

2. Digital4Better, digital4better/dataEvolving catalogue; pin a commit for reproducible factors. ↩

3. Uptime Institute historical PUE discussion ↩

4. Cloud Carbon Footprint Online documentation, unpinned version. ↩

5. Digital4Better, D4B commit 77cae4a. Snapshot 77cae4a, 2025 data; reviewed 16 October 2026. ↩