Glossary
Definitions reflect how terms are used in this methodology. Links lead to the calculations, assumptions and sources in the relevant chapter.
AWARE (Available WAter REmaining)
A characterization method for potential water deprivation affecting humans and ecosystems. It weights consumed volumes by local water availability. Results use m³ world-equivalent, distinct from physical m³ consumed. Detailed explanation.1
Actual lifetime
A lifetime estimated from equipment still in use and retired equipment. The final lifetime of equipment still operating remains unknown. See the estimator and its limitations.
Allocation
A rule assigning a study object its share of a shared resource's impacts. Main approaches include time, transferred volume, reserved capacity, user count and documented fixed shares. See Allocation and sharing.
Native application
An application running in a device’s operating-system environment, on mobile or desktop. This chapter also covers installed hybrid and cross-platform applications: distribution, local activity and remote services are inventoried by the resources actually used. See Native applications.
Hybrid application
An application combining native functions and embedded web content. Contributions are allocated without adding two complete inventories of the same journey. See the application boundary.
Background activity
Activity without direct interaction with the interface: synchronisation, notifications or deferred processing. Its duration is not equivalent to screen-on time; execution depends on the system and its restrictions. See platform specificities.
Archetype
A documented representative configuration for an equipment or component family. It provides a reference when actual characteristics are missing. Borrowed values remain estimates; suitability depends on range, generation, features and use context. See missing data.
Attributional approach
An approach that assigns a system’s impacts to the products or services studied. A smaller allocated share does not by itself demonstrate physically avoided emissions. Detailed explanation.
Reference barycentre
Geographical point representing a country in the reference dataset. The method uses the great-circle distance between these points to estimate distance between countries. It identifies neither a router nor a data centre and does not describe routing. See the network model.
Batch
A group of inputs processed together by a model. Batch size affects memory requirements and performance; its effect depends on hardware and the execution engine. Detailed explanation.
Boundary (system boundary)
The activities, resources and life-cycle stages included in a study, with its period and exclusions. An outsourced service can remain within the boundary when it contributes to the studied function. Detailed explanation.
CDN (Content Delivery Network)
A geographically distributed server network that delivers content near users. Caching can reduce origin-server traffic and transfer distances, while the CDN's own infrastructure remains within the relevant scope.
Characterization factor
A coefficient relating an inventory flow, such as an emission, to an impact indicator. It depends on the characterization method and version; it is not necessarily a whole-equipment factor. Detailed explanation.
Cradle-to-gate
Boundary from raw-material extraction to the product leaving the declared production stage. Downstream use and end of life are excluded; transport, packaging and infrastructure depend on the stated boundary. See data quality.
Double counting
Counting the same physical impact more than once through overlapping boundaries. See Double counting.
Electricity mix
The distribution of electricity generation sources for an area and period. Associated impact factors depend on technologies and inventories; a renewable share is not a zero-impact factor. Detailed explanation.
Embedding
A numerical vector representation of content, used notably to retrieve related content. Its production, storage and retrieval belong within a RAG application’s boundary. Detailed explanation.
Embodied impact
Environmental impacts of creating and disposing of hardware, including extraction, manufacturing, transport and end of life. The share assigned to a use depends on the allocation rule.
Endpoint approach
An indicator of potential damages, for example to human health or ecosystems. Moving from midpoints to endpoints requires additional models. Grouping indicators by theme alone is not an endpoint calculation.2
Environmental footprint
An assessment of pressures associated with human activities on resources and ecosystems, expressed here through multiple midpoint indicators.
Environmental indicator
A quantity describing an environmental effect or pressure after characterization. The nine selected indicators are distinct and cannot be added together. Their units and methods determine factor compatibility.
| Acronym | Meaning |
|---|---|
| GWP | Global warming potential |
| AP | Acidification potential |
| CTUe | Freshwater ecotoxicity |
| CTUh-nc | Human toxicity, non-cancer effects |
| CTUh-c | Human toxicity, cancer effects |
| PM | Health effects of particulate matter |
| IR | Effects of ionizing radiation |
| WU | Characterized impact of water use |
| ADPe | Abiotic depletion of mineral and metal resources |
Equipment
Electronic devices within the three digital tiers: user devices, network equipment and datacenter equipment. Used to display, process, store or transmit data.
Fine-tuning
Further training of an already trained model to adapt it to data or a task. The number of modified parameters alone does not determine computation cost. Detailed explanation.
Fixed-share allocation
A conventional fixed share assigned to a service when measured usage data are unavailable. Used as a documented fallback.
FLOP and FLOP/s
A FLOP is a floating-point operation; FLOP/s is a rate of operations per second. Workload and throughput are different quantities whose ratio can estimate a duration. Detailed explanation.
Functional unit and analysis unit
The functional unit describes the service delivered, its quantity and quality conditions. The analysis unit expresses the result, for example per session. A shared counting unit does not guarantee equivalent functions. Detailed explanation.
GPU (Graphics Processing Unit)
A processor suited to parallel computation, notably graphics and AI. Its assessment covers the included hardware life-cycle stages, operation and its share of server resources. Detailed explanation.
GPU-hour
One hour on one GPU; multiple GPUs over the same duration accumulate GPU-hours. A per-GPU-hour factor specifies whether its basis is active, reserved or calendar time, and which life-cycle stages it includes. See the AI use case.
HBM (High Bandwidth Memory)
High-bandwidth memory associated with some accelerators. Its manufacture and operation belong to the hardware boundary; HBM already included in a GPU factor is not added separately. Detailed explanation.
Hypervisor
Software that creates and manages VMs on a physical server, such as VMware ESXi, KVM, Hyper-V or Xen. It uses hardware resources and can introduce overhead.
Impact factor
Impact per reference quantity, such as a device, kWh, GB or duration. Its use requires units, life-cycle stages and a characterization method compatible with the study. Detailed explanation.
Inference
Execution of a trained model to produce an output from an input. Depending on the task, it includes prompt or media processing and response generation. Detailed explanation.
KV cache (Key-Value)
Retention of attention keys and values for previously processed tokens for reuse during generation. It avoids some recomputation but uses memory. Detailed explanation.
LCA (Life Cycle Assessment)
A standardized methodology, ISO 14040/14044, for assessing environmental impacts across a product or service life cycle: raw-material extraction, manufacturing, transport, use and end of life. Its principles also inform organizational assessment.
Life-cycle inventory
Inventory of flows needed for the study: materials, energy, emissions and waste within the retained boundary. It forms the basis for impact calculation and extends beyond an equipment list. Detailed explanation.
LoRA (Low-Rank Adaptation)
An adaptation method that trains low-rank matrices while keeping base weights frozen. Fewer trainable parameters do not automatically imply the same reduction in energy or operations. Detailed explanation.
Material criticality
Assessment of a material’s economic importance and supply risk. It is not an environmental impact indicator and differs from ADPe. Detailed explanation.
MFU (Model FLOPs Utilization)
The ratio of model-useful computational throughput to theoretical accelerator throughput under the retained conventions. It helps estimate duration; it is neither GPU activity nor a fraction of power consumption. Detailed explanation.
Midpoint approach
An indicator of an intermediate environmental problem, such as climate change in kg CO₂ eq. or acidification. It characterizes emissions and resource withdrawals without estimating final damages.2
Model parameters
Total parameters are stored weights; active parameters participate in the current computation. Trainable parameters can have their values updated. A frozen parameter keeps its value but can still participate in computation, including gradient propagation to other layers. MoE and weight freezing therefore affect different dimensions. Detailed explanation.
MoE (Mixture of Experts)
An architecture that activates only some model experts for each token. Active parameters contribute to computation; stored parameters remain relevant to memory estimation. Detailed explanation.
Nominal lifetime
An expected or conventional reference lifetime for equipment. It supports allocation when a better documented duration is unavailable; it is not a guaranteed minimum lifetime. See time and aging.
Operational impact
Use-phase impact, modeled here primarily from the electricity required to operate equipment.
PAIA (Product Attributes to Impact Algorithm)
Streamlined approach for estimating equipment footprints from product attributes. It identifies major contributions; its results and uncertainty alone are insufficient for ranking similar models. See manufacturer report interpretation.
PCF (Product Carbon Footprint)
A product footprint limited to climate impact. Included stages, configuration and usage scenario are specific to each report; the acronym guarantees neither a complete life cycle nor independent verification. See data quality.
Prefix cache
Reuse across requests of states computed for a shared prompt prefix. It reduces prefix processing without removing response computation or cache resource requirements. Detailed explanation.
Primary data
Data collected directly within the study scope or provided by an actor directly involved. Secondary data come from literature or generic reference datasets.
Prompt
Input supplied to a model: instructions, context and any media. Processing it and generating the response are distinct activities in the computation model. Detailed explanation.
Proxy
A substitute input or model used when direct information is unavailable. Its boundary, representativeness and limitations remain explicit; using it does not turn an estimate into a measurement. Detailed explanation.
PUE (Power Usage Effectiveness)
Total datacenter energy divided by IT-equipment energy. A PUE of 1.0 means no additional site energy beyond IT equipment, which includes compute, storage and networking. The value 1.58 is a historical fallback scenario, not a verified current global average.
Quantization
Representation of model values at reduced numerical precision. It can change memory requirements, throughput and output quality; it does not automatically reduce the number of parameters. Detailed explanation.
RAG (Retrieval-Augmented Generation)
Generation augmented with content retrieved from a document collection. Assessment also covers embeddings, indexing, retrieval and storage, beyond the model call alone. Detailed explanation.
Equivalent radius
Radius of a disk with the same area as the country: , with in km² and in km. It estimates internal distance in the network model. The associated circle follows neither borders nor network coverage. See the network model.
RCP (Product category rules)
Environmental assessment rules for a product or service category. Sector-specific rules extend the parent digital-services PCR. Their use here does not establish overall compliance. See D4B’s positioning.
REF and CFE
REF describes renewable energy share. CFE describes carbon-free energy under the provider's convention and may include nuclear energy. These percentages are not interchangeable and do not automatically reduce a location-based electricity factor.
Reserved capacity
Resources made available to a service, even when little used. Reservation-based allocation assigns that committed capacity; it does not directly measure actual consumption. Detailed explanation.
Resource sharing
Sharing hardware between services or users, including through virtualization or containers. It can reduce physical equipment requirements when capacity is actually consolidated.
Robustness matrix
A D4B documentation grid across six dimensions: boundary, geography, time, physical description, traceability and method. Each dimension receives a justified score from 1 to 5. It is not certification. See data quality.
Robustness score
The mean of the six documentation ratings in the D4B grid. Individual dimensions and blocking incompatibilities take precedence over that mean; it is neither a probability of correctness nor a confidence interval. See data quality.
SCI and SCI-AI
SCI (Software Carbon Intensity) describes software carbon intensity per functional unit. SCI-AI specifies its application to AI. These carbon frameworks do not replace a multi-impact assessment. Detailed explanation.
Secondary data
Data from literature, a database or a context other than direct measurement of the studied system. Suitability depends on compatibility and representativeness, not only age. Detailed explanation.
Sensitivity analysis
Variation of assumptions within justified bounds to observe their effect on the result. It identifies inputs that influence the conclusion; it does not provide a statistical confidence interval. Detailed explanation.
TDP (Thermal Design Power)
A component’s thermal design power under the manufacturer’s conventions. It is not a measurement of average server power or energy consumed over a period. Detailed explanation.
TEC (Typical Energy Consumption)
Typical electricity consumption associated with a protocol and operating profile. Annual TEC includes that profile’s modes and durations; it is not measured active power for an application. See the Native applications.
Third-party service
A service supplied by an external actor and used by the studied function, such as an API, payment, video or cloud service. Outsourcing does not justify exclusion; overlaps with other contributions still need accounting. Detailed explanation.
Time allocation
Assigning impacts in proportion to the service's usage duration relative to equipment lifetime. A default rule for dedicated devices and servers.
Token
A unit produced by a model’s text tokenization. A token is not equivalent to a word: counts depend on model, language and content. Media representations follow their own conventions. Detailed explanation.
Training
Adjustment of a model’s parameters using data. Processed volume counts every presentation of the data; experiments and preparation can also contribute to the model life cycle. Detailed explanation.
User device
Electronic equipment through which a user interacts with a computing system: computers, smartphones, tablets, payment terminals, ATMs and others.
vCPU
A virtual processing unit assigned to a VM. Its relationship to physical threads and cores depends on the offering, architecture, simultaneous multithreading and oversubscription. That relationship is part of the allocation assumptions.
Versioning
Recording the versions of data, factors and methods used in an assessment to support reproducibility and comparison over time.
VM (virtual machine)
An execution environment using a share of a physical host’s resources. It is not additional hardware; its impacts derive from allocation of host resources and associated infrastructure. Detailed explanation.
Volume allocation
Assigning impacts in proportion to transferred data volume. This attribution convention does not demonstrate physical consumption proportional to bytes. See the network model.
WUE (Water Usage Effectiveness)
Direct site water use divided by IT-equipment energy, usually in liters per kWh IT. Its definition identifies whether the numerator represents withdrawal or consumption, along with the period and site. Distinct from WU, a characterized water-impact indicator.
Reference configuration
An explicit component assembly used when an equipment’s exact composition is unknown. Characteristics are documented assumptions; impacts are recalculated with component models. A configuration is not a renamed flat impact factor.
Calibration
Estimating or adjusting model parameters using independent, compatible evidence. Validation then assesses its ability to reproduce observations distinct from the fitting data.
Derived factor
A model result normalised to a unit, such as device manufacturing per second. Reproduction requires the configuration, coefficients, boundaries and allocation convention.
LAN
Local network connecting devices on a site: switches, access points, cabling and supporting equipment. Its boundary differs from Internet access.
GB·km
A volume in gigabytes multiplied by a distance in kilometres. In the D4B network model, 2 GB over a conventional distance of 100 km represents 200 GB·km. This is an allocation-model quantity, not a direct energy measurement. See Network.
Codebook
A learned set of codes representing a signal in discrete form. An audio model can combine several codebooks to represent different levels of detail; their count alone does not determine actual compute cost. Detailed explanation.
Diffusion (generative model)
Generation through successive denoising steps, guided by instructions or media. DiT uses a Transformer; U-Net uses an architecture with downsampling followed by upsampling. Step count, processed positions and architecture determine the workload. Detailed explanation.
Latent space
An internal media representation, often more compact than pixels or audio samples. Computation can operate on these latent positions. A VAE (variational autoencoder) can map between this representation and the media; decoding is a processing stage to inventory. Detailed explanation.