AI Datacenter Energy & Efficiency Glossary

Comprehensive reference for metrics, cooling, power systems, carbon accounting, and AI/HPC-specific terminology in modern datacenters.

A

Absorption chiller

A heat-driven chiller that uses a thermal source (e.g., waste heat or steam) instead of an electric compressor to produce chilled water. Can support heat-reuse strategies.

Cooling

AC–DC conversion losses

Energy lost (typically a few percent per stage) when rectifying AC utility power to DC for IT loads (inside PSUs, UPS rectifiers, etc.). Lowering the number of conversions reduces losses.

Power

Active power (kW)

The portion of electrical power that performs useful work; contrast with apparent power (kVA) and reactive power (kVAR).

Power

Adiabatic cooling

Pre-cooling of intake air by evaporating water upstream of a dry cooler or condenser. Reduces compressor hours; increases site water use (see WUE).

Cooling

Advanced metering (power/thermal)

Granular, timestamped measurements (rack-level or device-level) for power, temperatures, flow, pressures; foundation for PUE, RTI, and energy analytics.

Monitoring

AI accelerator

Specialized processor (GPU, TPU, NPU) optimized for matrix/tensor operations used in AI training and inference; typically the dominant driver of rack and hall power density.

AI/HPC

AI inference energy per token (J/token)

Energy divided by number of output tokens generated; useful for service-level energy benchmarking of LLMs. See also Performance per watt.

AI/HPC

AI training run energy (MWh per run)

Total electrical energy to complete a training job, including accelerators, CPUs, networking, and cooling; often converted to kgCO₂e per run via grid emission factors.

AI/HPC

Allowable vs. Recommended environmental envelopes (ASHRAE)

"Recommended" is the narrow band for steadier reliability/efficiency; "Allowable" is the wider, tested safe band. Classes A1–A4 define equipment tolerance to temperature/humidity. Operating closer to the warm side of "Recommended" increases economizer hours and reduces compressor power.

Standards

Apparent power (kVA)

Product of RMS voltage and current in AC systems; includes both active and reactive components. See Power factor.

Power

Approach temperature (heat exchanger)

The difference between the leaving fluid temperature of a heat source and the entering fluid temperature of the sink. Lower approach → better heat transfer → higher efficiency.

Cooling

ATS (Automatic Transfer Switch)

Switchgear that transfers load between utility and generators (or other sources) automatically on power disturbance.

Power

B

Battery Energy Storage System (BESS)

A utility-scale battery (often Li-ion) at the site for peak shaving, grid services, backup bridging, or renewables integration; complements but does not replace the UPS ride-through function. Typical energy duration is 1–2 hours in data centers.

Power

Blowdown (cooling tower)

The controlled discharge of concentrated cooling water to limit mineral buildup. Managed together with Cycles of concentration (CoC) to reduce water use and scaling.

Cooling

Busway (overhead bus)

Modular, overhead power distribution with plug-in tap-offs to racks, reducing underfloor whips and improving reconfiguration and losses at scale.

Power

Bypass airflow (cooling)

Cold air supplied but not ingested by IT equipment (e.g., leaks around racks); wastes fan and chiller energy. Mitigated by sealing and containment.

Cooling

C

Carbon-free energy (CFE), 24/7 CFE

Electricity matched hour-by-hour with carbon-free generation (renewables and nuclear) on the same grid where consumption occurs—stronger than annual REC matching. See Granular certificates (GCs).

Carbon

Carbon intensity (gCO₂e/kWh)

Emissions per unit of electricity; varies hourly by grid mix. Used to convert kWh to emissions (Scopes 2 & sometimes 3 with upstream losses).

Carbon

Carbon Usage Effectiveness (CUE)

A standardized metric of total datacenter CO₂e emissions divided by IT energy (kgCO₂e/kWh_IT). Complements PUE by incorporating carbon.

Metrics

CDU (Coolant Distribution Unit)

A heat-exchange and pumping skid that interfaces facility water and IT-loop coolant for direct-to-chip or immersion systems; provides monitoring and leak containment.

Cooling

Chiller COP / kW/ton

Coefficient of Performance (cooling kW produced per kW input). In Imperial units, efficiency is often expressed as kW/ton (lower is better); 1 ton ≈ 3.517 kW of cooling.

Cooling

CIP (Critical Infrastructure Power)

The portion of site capacity allocated to essential plant (UPS, chillers, pumps, CRAH fans).

Power

Cold aisle / hot aisle containment

Physical barriers that prevent mixing of supply and exhaust air, raising return temperature and efficiency; hot-aisle containment typically maximizes economizer hours.

Cooling

Cold plate (direct liquid cooling)

A metal plate bonded to chips with coolant flowing through internal channels to remove heat at high flux.

Cooling

Cooling Efficiency Ratio (CER)

ISO/IEC 30134-7 metric relating cooling capacity delivered to energy consumed by cooling infrastructure. See also PUE.

Metrics

Cooling tower

An evaporative heat rejection device in water-cooled plants; performance and water use governed by CoC and ambient wet-bulb conditions.

Cooling

CoP (Coefficient of Performance)

Ratio of useful thermal output to electrical input for chillers/heat pumps; a dimensionless efficiency figure.

Cooling

Cycles of concentration (CoC)

Ratio of dissolved solids in recirculating tower water to those in make-up water; higher CoC reduces make-up and blowdown volumes (within corrosion/scaling limits).

Cooling

CRAC / CRAH

Room air conditioners/handlers. CRAC uses direct expansion (DX) refrigeration; CRAH uses chilled water coils and fans. Both deliver conditioned air to IT intakes.

Cooling

Critical load

The IT and essential support load that must remain powered during disturbances.

Power

D

Data Center infrastructure efficiency (DCiE)

The reciprocal of PUE (IT energy divided by total facility energy). DCiE = 1/PUE.

Metrics

Data hall density (kW/rack)

Average or planned IT power per rack; AI training clusters commonly require 20–80+ kW/rack (air) and 80–140+ kW/rack (liquid).

Design

DCIM / EPMS

Datacenter Infrastructure Management (capacity, assets, thermal) and Electrical Power Monitoring System (real-time power/energy/quality) platforms for efficiency and risk control.

Monitoring

DER (Distributed Energy Resources)

On-site PV, fuel cells, BESS, generators, and controllable loads used for reliability and tariff optimization; enables grid services.

Power

Direct-to-chip (D2C) liquid cooling

Liquid circulated through cold plates directly attached to high-power devices (CPUs/GPUs/ASICs). Removes most heat as liquid, dramatically reducing fan and compressor work.

Cooling

Direct expansion (DX)

Refrigerant-based cooling where refrigerant evaporates directly in the coil; common in CRACs and some modular systems.

Cooling

Double-conversion UPS (VFI, online)

UPS topology where the rectifier and inverter run continuously, isolating the load from input disturbances; highest protection with some conversion losses.

Power

Dry cooler

Air-cooled heat exchanger rejecting heat without evaporation; enables water-free or winter "free-cooling" operations.

Cooling

DVFS (Dynamic Voltage and Frequency Scaling)

Silicon power-management technique adjusting voltage/frequency to match workload; saves energy when utilization is low or latency bounds allow.

AI/HPC

E

Economizer (air-side / water-side)

Uses favorable ambient conditions to provide cooling with reduced or no compressor operation; key to seasonal efficiency gains.

Cooling

Energy Reuse Effectiveness (ERE)

Metric that credits beneficial reuse of datacenter heat (e.g., space heating, process heat). ERE modifies PUE by subtracting reused energy from facility energy.

Metrics

Energy Reuse Factor (ERF)

Fraction of total datacenter energy that is beneficially reused outside the datacenter. Standardized in ISO/IEC 30134-6.

Metrics

Energy-proportional computing

System that scales power closely with workload (near-zero idle overhead), minimizing energy at low utilization—important for inference fleets.

AI/HPC

EAC / REC / GO

Energy Attribute Certificates. In North America, RECs; in Europe, Guarantees of Origin (GOs). Each certificate typically represents the environmental attributes of 1 MWh of renewable generation; used for market-based Scope 2 accounting and claims.

Carbon

EER / IPLV / NPLV

Cooling performance ratings. EER = full-load efficiency; IPLV/NPLV = integrated part-load performance based on weighting curves.

Cooling

F

Fan energy index / pressure

Fan power scales roughly with the cube of airflow; better containment and lower pressure drops reduce fan energy.

Cooling

Free cooling hours

Annual hours when outdoor conditions allow economizer operation; depends on server inlet setpoints per ASHRAE classes.

Cooling

Frequency regulation (grid service)

Rapid up/down power response to stabilize system frequency; datacenters can provide this via grid-interactive UPS and flexible loads.

Power

Fuel cell (SOFC/PEM)

On-site generation converting fuel's chemical energy to electricity with high efficiency and low local emissions; can be configured for combined heat and power (CHP).

Power

G

Granular certificates (GCs), hourly certificates

Certificates time-stamped at hourly (or sub-hourly) granularity to verify 24/7 CFE matching, advancing beyond annual REC matching. See EnergyTag.

Carbon

Green500 (FLOPS/W)

Supercomputing ranking by energy efficiency (GFLOPS per Watt on LINPACK); a proxy for accelerator platform efficiency.

AI/HPC

Grid-interactive efficient building (GEB)

Facility that flexes load, storage, and generation to support the grid, often monetizing demand response while maintaining service-level constraints.

Power

H

HBM (High-Bandwidth Memory)

Stacked DRAM adjacent to accelerators; significant share of board power in AI/HPC servers and often a thermal constraint.

AI/HPC

Heat pump for heat reuse

Upgrades low-grade waste heat from the datacenter (e.g., 30–45 °C liquid return) to higher temperatures suitable for district heating.

Cooling

Heat Reuse

Export of recovered thermal energy to adjacent buildings/district networks; credited by ERF/ERE.

Cooling

Hot-aisle recirculation

Hot exhaust that returns to IT intakes; increases IT inlet temperatures; mitigated by containment and blanking/sealing.

Cooling

I

IEC 62040-3 (VFI/VI/VFD)

UPS performance classification; VFI ≈ online/double-conversion; VI ≈ line-interactive; VFD ≈ standby/offline.

Standards

Immersion cooling (single-phase / two-phase)

IT submerged in dielectric fluid; single-phase uses pumped liquid, two-phase uses boiling/condensation for very high heat flux. Cuts fan power and can simplify heat reuse.

Cooling

Indirect evaporative cooling (IEC)

Uses evaporation on a secondary air stream or fluid to cool supply air without adding moisture to the white space.

Cooling

In-row cooling

Cooling units between racks that shorten airflow paths and improve return temperature.

Cooling

ITEEsv / ITEUsv (ISO/IEC 30134-4/5)

Server-level energy efficiency/utilization metrics standardized for comparability across IT equipment.

Metrics

J

Joules per token

See AI inference energy per token.

AI/HPC

K

kW/ton (chillers)

Electrical kW input per "ton" of cooling (3.517 kW); lower is better. Useful for whole-plant efficiency tracking.

Cooling

L

Leak detection system (LDS)

Fiber or point sensors in DLC/immersion loops and under raised floors to quickly detect leaks.

Monitoring

Life-cycle assessment (LCA)

Accounting of embodied and operational impacts (Scope 3 + Scope 1/2) over datacenter assets' full life.

Carbon

Liquid-to-liquid (L2L) / Liquid-to-air (L2A)

Heat exchanger coupling (e.g., rack loop to facility loop for L2L, or rack rear-door to room air for L2A).

Cooling

Load factor (energy)

Ratio of average demand to peak demand over a period; higher load factor improves asset utilization and often reduces tariffs.

Power

M

Marginal emissions rate

The incremental grid emissions from a small change in load; useful for scheduling flexible AI workloads to low-carbon hours.

Carbon

Microgrid

On-site generation + storage + controls that can island from the grid; supports resilience and CFE goals.

Power

MIG / Multi-Instance GPU

Partitioning a GPU into smaller instances to increase utilization; can improve fleet-wide energy efficiency per request.

AI/HPC

N

Nameplate vs. measured power

IT gear often consumes far less than nameplate at typical utilization; plan power based on measured profiles and realistic diversity.

Power

N+1 / 2N (redundancy)

Design reserve for power/cooling paths. Higher redundancy adds losses (more equipment, part-load operation) but improves availability.

Design

O

Open Compute Project (OCP) liquid cooling

Open standards for manifolds, CDUs, and components to accelerate DLC adoption and interoperability.

Standards

P

Partial PUE (pPUE)

PUE measured for a subsection (e.g., a single room or container) to isolate efficiency impacts of a subsystem.

Metrics

Peak shaving / load shifting

Using BESS or control to reduce demand charges or move compute to lower-cost/low-carbon hours.

Power

Performance per watt

Compute work per unit power (e.g., GFLOPS/W). Core efficiency comparator for AI/HPC hardware and clusters.

AI/HPC

Power capping

Enforcing platform-level limits (server/GPU) to keep racks within power and thermal envelopes; can reduce J/token with minimal throughput loss in some scenarios.

Power

Power factor (PF)

Ratio of real power (kW) to apparent power (kVA). PF close to 1 reduces conductor/transformer losses and utility penalties.

Power

Power quality (THD/TDD)

Harmonic distortion and related power anomalies that cause extra heating/losses and stress. Managed per IEEE-519 guidelines with filters and proper equipment sizing.

Power

PPA (Power Purchase Agreement)

Contract to buy power (and often attributes) from off-site generation; virtual/sleeved structures are common for hyperscalers; for carbon claims see Scope 2.

Carbon

PUE (Power Usage Effectiveness)

The foundational datacenter efficiency metric: PUE = Total facility energy ÷ IT energy. 1.0 is ideal. Standardized in ISO/IEC 30134-2.

Metrics

Pump energy (kW)

Energy for chilled-water and condenser-water pumps or rack-loop pumps. Variable-speed drives and optimized ΔT/ΔP control cut kWh.

Power

R

Rack Cooling Index (RCI)

Percentage-based measure of how well rack inlet temperatures meet ASHRAE bands (HI/LO versions); complements RTI.

Metrics

RDP (Rear Door Heat Exchanger / RDHx)

Liquid-cooled heat exchanger mounted at rack rear to capture and remove heat at the source, enabling higher rack densities without full DLC.

Cooling

REF (Renewable Energy Factor)

ISO/IEC 30134-3 metric expressing the proportion of a datacenter's energy sourced from renewables.

Metrics

Return Temperature Index (RTI)

Ratio comparing rack ΔT to air-handler ΔT; indicates bypass (RTI < 100%) or recirculation (RTI > 100%). Used to diagnose airflow efficiency.

Metrics

S

Scope 1 / 2 / 3 (GHG Protocol)

Scope 1 = direct onsite emissions; Scope 2 = purchased energy (location-based and market-based reporting); Scope 3 = value-chain (e.g., embodied IT hardware).

Carbon

SEER / Seasonal metrics

Seasonalized efficiency indices that reflect part-load operation across outdoor conditions; helpful for comparing cooling options in specific climates.

Cooling

Server inlet setpoint

Target supply temperature at IT intakes; raising setpoints within ASHRAE "Recommended" increases economizer use and plant efficiency.

Cooling

SHI/RHI (Supply/Return Heat Index)

Dimensionless indices assessing thermal separation quality between supply and return airstreams; values near 1 (or 100%) indicate minimal mixing.

Metrics

Single-line diagram (SLD)

Electrical one-line showing sources, distribution, protection, and critical loads; used for loss analysis and energy metering strategy.

Design

Site energy vs. Source energy

Site kWh is what you meter; source kWh accounts for upstream generation/transmission losses; used in some sustainability reporting.

Carbon

T

Thermal energy storage (TES)

Storing cooling (e.g., chilled water or phase-change media) to shift compressor work off peak and support grid interaction.

Cooling

Total Usage Effectiveness (TUE)

Comprehensive figure of merit proposed for chip-to-facility energy pathways; decomposes losses across IT, power, and cooling subsystems.

Metrics

Transformer losses (no-load / load)

Core and copper losses that vary with loading. Right-sizing and high-efficiency models (e.g., DOE 2016) reduce wasted kWh.

Power

Two-phase DLC (boiling)

Uses fluid phase change at the cold plate for very high heat flux removal; requires careful controls and compatibility.

Cooling

TVD / Thermal design power (TDP/TBP)

Nameplate thermal power for a chip or accelerator board to be removed by cooling. Real workloads may spike above TDP.

AI/HPC

U

Uptime Institute Tiers (I–IV)

Topology standards describing redundancy/fault tolerance of the MEP design; higher tiers typically incur more idle losses unless carefully optimized.

Standards

UPS eco-mode / high-efficiency modes

Bypass-dominant UPS modes raising efficiency at the expense of some conditioning; used selectively based on power quality and SLAs.

Power

Utilization (IT)

Fraction of peak compute capacity being used. Increasing average utilization often yields the largest practical efficiency gains (same energy amortized over more work).

AI/HPC

V

Variable frequency drives (VFDs)

Motor drives that modulate speed for pumps and fans; energy savings scale steeply with reduced speed.

Power

Virtual PPA (vPPA)

Financial contract for differences linked to a renewable project; supplies market-based Scope 2 benefits without physical delivery.

Carbon

W

Water Usage Effectiveness (WUE)

Total site water use per unit of IT energy (L/kWh_IT). Complements PUE by addressing water footprint; standardized in ISO/IEC 30134-9 and originally defined by The Green Grid.

Metrics

Water-side economizer

Uses low ambient temperatures to cool water via dry coolers or cooling towers with minimal chiller operation.

Cooling

Wet-bulb temperature

Governs evaporative cooling effectiveness; lower wet-bulb increases free-cooling hours and reduces compressor energy.

Cooling

Workload placement (carbon-aware)

Scheduling AI jobs to regions/hours with lower marginal emissions or high CFE; reduces CUE without hardware changes.

Carbon

Quick Formulas

Key Metrics

  • PUE = Total Facility Energy ÷ IT Energy (ideal → 1.0)
  • DCiE = IT Energy ÷ Total Facility Energy (= 1/PUE)
  • WUE = Annual Site Water (L) ÷ Annual IT Energy (kWh_IT)
  • CUE = Annual Data Center CO₂e (kg) ÷ Annual IT Energy (kWh_IT)

Performance Indicators

  • RTI (%) ≈ (Rack ΔT ÷ Air-handler ΔT) × 100
  • <100% = bypass, >100% = recirculation
  • GFLOPS/W (Green500) = Performance ÷ Power

Standards & Programs

ISO/IEC 30134 Series

Data centre metrics: PUE (-2), REF (-3), ITEE/ITEU (-4/-5), CER (-7), CUE (-8), WUE (-9)

ASHRAE TC 9.9

Thermal Guidelines (A1–A4 environmental classes; allowable vs recommended)

Other Key Standards

  • • Uptime Institute Tiers (I–IV) for topology/resilience
  • • EU Code of Conduct for Data Centre Energy Efficiency
  • • GHG Protocol Scope 2 (location-based and market-based accounting)
  • • EnergyTag (granular hourly certificates for 24/7 CFE)