Processors and accelerators

This chapter covers CPUs, GPUs, NPUs, TPUs and other specialised accelerators. It addresses embodied impacts, operational power and completion rules where specifications are incomplete. The model relies on observed or documented characteristics when available and clearly identified archetypes otherwise.

Characteristics

Characteristic	Unit	Treatment	Purpose
Component family	Category	Documented archetype	CPU, GPU or specialised accelerator
Model / reference	Identifier	Unknown if unavailable	Map to suitable specifications and inventories
Compute units / cores	Count	Model or archetype	Capacity description; not a universal impact factor
TDP / rated power	W	Documented specification	Scenario input, distinct from measured average power
Associated memory	GB	Exact configuration where possible	Material for accelerators; HBM counted once
Die area	mm²	Unknown if unavailable	Semiconductor model input; package area is not interchangeable
Component count	Units	Actual count	Scale to equipment configuration

Modelling principle

A component model can distinguish configuration and generation where its inputs are supported. A representative aggregate inventory may be preferable when detailed reconstruction is speculative.¹

Available evidence	Model representation
Precisely identified component	Observed or documented model characteristics.
Partly identified component	Compatible archetype from the same family.
Unknown component, known host	Inference from equipment type and use context.

See Semiconductors and Equipment modelling.

Embodied impacts

Factor status

Historical GWP, ADP and primary-energy coefficients are documented by Boavizta from Green Cloud Computing. Earlier additional AP, WU, PM and IR coefficients described as an EF reconstruction are not retained: the repository does not establish the elementary flows, compartments, characterisation factors and reproducible calculation needed to support them. A logic-CPU inventory cannot simply be transferred to DRAM or NAND.

Missing factors are data gaps, never zero impacts. Technology and packaging multipliers remain educational scenarios until calibrated. The source's ADP label does not prove equivalence to another dataset's ADPe.

Parametric formula

Die area

Silicon area is a key input to the manufacturing model.

Technology node

The process determines manufacturing steps, energy and water requirements.

Packaging

Assembly, including advanced packaging, contributes to manufacturing impacts.

Technology and packaging multipliers remain scenario assumptions until calibrated. The exploratory model is:

$$I_{emb}=(S_{die}C_{die}+C_{base})F_{node}F_{pkg}N_{units}$$

$S_{die}$ is in mm²; $C_{die}$ is impact per mm²; $C_{base}$ is the historical fixed contribution; $F_{node}$ and $F_{pkg}$ are scenario multipliers; $N_{units}$ is component count.

Indicator	Die coefficient per mm²	Fixed base	Unit
GWP	1.97E-02	9.14	kg CO₂ eq.
Historical ADP	5.87E-09	2.04E-02	kg Sb eq.
Primary energy	0.265	156.0	MJ
Toxicity indicators	No validated coefficient here	;	Compatible inventory required

Source: Boavizta CPU methodology², implementing historical Green Cloud Computing coefficients. Node and packaging multipliers below are D4B assumptions, not a verified imec coefficient table.

A multicriteria reconstruction requires a complete inventory, identified LCIA method and reproducible characterisation. It is a research direction, not existing indicator coverage.

Exploratory node factors

Scenario node	$F_{node}$	Status
28 nm	1.0	Educational reference
22 nm	1.05	D4B assumption
14 nm	1.2	D4B assumption
10 nm	1.4	D4B assumption
7 nm	1.8	D4B assumption
5 nm	2.2	D4B assumption
4 nm	2.5	D4B assumption
3 nm	2.8	D4B assumption

Node names do not fully describe a foundry process. These multipliers establish no universal mapping to EUV, FinFET or GAA. Likewise, a packaging multiplier of 1.5 for an accelerator is a scenario, not an inventory of advanced packaging.

Ex: an 800 mm² scenario with multipliers 2.2 and 1.5 gives $(800\times0.0197+9.14)\times2.2\times1.5=82.17$ kg CO₂ eq. This is not a commercial GPU LCA. The fixed term also means total impact does not scale exactly with die-area ratios.

Manufacturing geography and electricity

The electricity mix at the manufacturing location affects the impacts of electricity consumed by the processes. Each process location must come from the component’s supply chain. Countries in the assumptions register remain scenarios when this information is missing.

Example: 100 kWh of manufacturing electricity

For a process consuming 100 kWh, two hypothetical locations give the following climate results using D4B’s 2025 average consumption factors.³

Scenario location	Electricity factor (kg CO₂e/kWh)	Impact for 100 kWh
France	0.0319271	3.19 kg CO₂e
China	0.589301	58.93 kg CO₂e

The calculation is energy × electricity factor. The 100 kWh is an illustrative assumption, not average consumption per chip. These results cover only electricity for this process, excluding materials, direct emissions and other processes.

Changing the country of use does not change the manufacturing location. Adapting a manufacturing inventory to another country requires separating its electricity consumption: the ratio between two electricity factors does not apply to the component’s total impact. The electricity impact calculation describes this calculation for each indicator. IR factor units must be verified before aggregation.

Reproducible base-model examples

With $F_{node}=F_{pkg}=1$:

Die area	Count	Calculation	Total embodied GWP
100 mm²	1	$100\times0.0197+9.14$	11.11 kg CO₂ eq.
200 mm²	1	$200\times0.0197+9.14$	13.08 kg CO₂ eq.
400 mm²	2	$(400\times0.0197+9.14)\times2$	34.04 kg CO₂ eq.

These illustrate the model rather than commercial products. A full graphics card also contains memory, board, cooling and other assemblies and lifecycle stages. The current Boavizta GPU methodology⁴ models a broader decomposition; a CPU multiplier is not a substitute for it.

Coverage

Indicator	Current treatment
GWP	Historical CPU model plus exploratory multipliers
Historical ADP	Source coefficients; aggregation requires compatibility with other methods
Primary energy	Historical model, distinct from final electricity
WU	No validated characterised component factor; physical water is not directly additive
AP, PM, IR, CTUe, CTUh	Compatible product LCA or inventory needed; whole-equipment factors do not prove a component split

When data are missing

The exact component reference is sought first. If unavailable, a characteristic may be borrowed from a neighbouring model in the same family, generation and range. This choice needs justification: matching core counts do not prove that two chips are equivalent.

Situation	Fictional example
Starting information	The CPU has 16 cores and a known generation. Its exact reference and silicon area are unknown.
Selected reference	A comparable CPU has a die area of 200 mm². This area is hypothetical in this example.
Calculation inputs	The formula uses an estimated 200 mm² die area. The 16 cores remain known data; they were not used to calculate that area.

If no neighbouring model is suitable, a documented family profile can be used, followed by a documented median or precautionary assumption as a last resort. Without justification, the input remains missing. A CPU profile cannot replace a GPU profile.

End of life

End-of-life coverage depends on the selected embodied inventory. It is not separately established for every processor here. Historical component factors do not necessarily share the same lifecycle boundary.

Operational impacts

Power profile

CPU

Power often varies between idle, intermediate and high-load operation.

GPU

Sustained computation can bring power closer to the rated value.

Specialised accelerators

Profiles may be steadier but remain strongly task-dependent.

Power varies by workload, scheduling and power state. Where no representative curve is available, the scenario identifies the rated-power proxy, load assumptions and operating duration. TDP is not a measured wall-power average.

Adjustments

The assessment covers utilisation, workload efficiency, host-resource sharing and site overhead. MFU relates model arithmetic work to theoretical throughput; it is not directly a fraction of TDP. PUE applies once at the correct IT/site boundary; quantities already containing it require no further multiplier.

Allocation and annualisation

Processors normally belong to a host device and support multiple workloads. Manufacturing allocation depends on retained lifetime, observation time and resource share. CPU-hours, GPU-hours or reserved capacity can provide suitable allocation quantities, but their meanings must remain consistent.

See Allocation and sharing.

Assumptions and limitations

Documented specifications can support a scenario; they do not validate every environmental coefficient. Representative archetypes are preferable to arbitrary values. Real power varies with software and load. Embedded memory, advanced packaging and cooling may be poorly isolated, and some inventories remain whole-equipment aggregates.

Further work should connect component specifications to process inventories, improve specialised-accelerator data, measure workload power profiles and calibrate the existing exploratory simulator.

Interactive example

Component type

Die size (mm²)

Process scenario (assumption)

Number of units

Usage conditions and lifetime

TDP / Nominal power (W)

Daily usage (hours)

Load factor (%)

Lifetime (years)

Country (electricity mix)

PUE

Results (per year)

Climate (GWP)

Embodied: 2.62 kgCO2eOperational: 44.22 kgCO2eTotal: 46.84 kgCO2e

Embodied

Operational

Operational electricity

1.39 MWh

Resources (historical ADP)

Embodied: 4.08 g Sb eq.Operational: 869.03 mg Sb eq.Total: 4.95 g Sb eq.

Embodied

Operational

CPU baseline GWP: area × 0.0197 + 9.14 kgCO₂e (Boavizta). Process and packaging adjustments are exploratory, not IMEC-validated. Manufacturing water is not calculated because physical water volumes cannot be added to characterized electricity water impacts.

Use-phase electricity: D4B snapshot, 2025 annual averages. The selected country applies to use; it does not relocate manufacturing factors.

Applications

Generative AI, cloud, and servers.

Footnotes

1. Versioned data/cloud/cpus.* and data/cloud/accelerators.* in Digital4Better data; Boavizta methods cited above; publications including Green Cloud Computing, GPU LCA material and the multicriteria A100 lifecycle study. Digital4Better data catalogue, snapshot 77cae4a. References and scope: Sources and corpus. ↩

2. Boavizta CPU methodology. Online documentation, unpinned version. ↩

3. Digital4Better, country electricity factors, year 2025, snapshot 77cae4a, GWP, FR / CN. ↩

4. Boavizta GPU methodology. Online documentation, unpinned version. ↩