Datacenter Glossary

A redundancy model in which every system component is fully duplicated, allowing one complete half to fail with no impact on the load.

An IEEE 802.3bs Ethernet standard at 400 Gb/s, the current spine-uplink standard for AI factory networks; 800GbE (IEEE 802.3df) follows for next-generation AI fabric deployments.

A large-scale GPU cluster purpose-built for AI model training and inference, requiring specialised high-density power, liquid cooling, and high-speed network fabric infrastructure.

Physical barriers preventing hot exhaust and cold supply air from mixing in a data hall.

ASHRAE TC9.9 defines classes A1–A5 and B/H for allowable IT equipment inlet air temperatures, guiding data centre cooling design.

Spare energy capacity available for additional IT workloads, expressed over a measurement period.

Spare power capacity available for additional IT load: the headroom between current IT draw and the rated IT load.

The average temperature of waste heat leaving the data centre: determines suitability for heat recovery applications.

Energy produced by backup generators during the measurement period: tests, mains outages or peak-shaving.

A physical server provisioned directly to a single workload without a virtualisation layer, delivering maximum compute performance and predictable latency at the cost of reduced density and flexibility.

Large-scale battery installations that store electrical energy for grid balancing, peak shaving or backup power, increasingly used alongside or instead of diesel generators.

A filler panel installed in unused rack unit spaces to prevent hot exhaust air from recirculating to the front of the rack.

Central control system for a building's mechanical and electrical services including HVAC, power, and access control.

The process of tracking, forecasting and planning available power, cooling, space and network resources in a data centre to meet future demand.

Annual CO₂ equivalent emissions from data centre energy divided by IT equipment energy in kWh.

A refrigeration machine that produces chilled water for distribution to CRAH units or CDUs in a data centre cooling system.

Service model where customers own their servers but rent space, power, and connectivity from a facility operator.

Standalone precision cooling unit with an internal compressor used to cool server halls.

Fan-coil unit that uses chilled water from a central plant to cool data hall air. More efficient than CRAC at scale.

A centralised repository of information about IT infrastructure components and their relationships, used to support change and incident management.

A facility unit that conditions and distributes liquid coolant to direct-liquid-cooled server racks, managing flow rate, supply temperature, and pressure between the facility chilled water loop and rack manifolds.

How efficiently the cooling system removes heat from the IT space relative to the cooling energy it consumes. Higher is better.

Total energy used by all cooling systems: CRAHs, CDUs, chillers, cooling towers, pumps and fans.

Cooling energy as a share of IT energy — kWh of cooling per kWh of IT load.

A heat rejection device that transfers waste heat from the condenser water loop to the atmosphere through evaporation.

NVIDIA's parallel computing platform and programming model that enables GPUs to be used for general-purpose computation, including AI and machine learning.

Inverse of PUE as a percentage. DCiE = (IT Power ÷ Total Power) × 100. Higher is better.

Software platform that monitors, measures, manages, and controls all data centre infrastructure and IT assets.

A programmable processor that offloads networking, storage and security functions from the CPU, freeing it for application workloads.

Total energy consumed by the data centre across all systems: IT, cooling, power distribution, lighting and ancillaries.

The temperature difference between the supply and return of a cooling circuit; a key indicator of how efficiently heat is being transferred.

A rotary UPS that integrates a flywheel, diesel engine and generator into a single unit, eliminating static battery banks.

Cooling method bringing liquid directly to heat-generating components, enabling high-density rack deployments.

Electricity used to raise waste heat to a usable temperature via heat pumps, enabling heat recovery from low-grade sources.

The European standard series for data centre facilities, covering power, cooling, physical security and management systems with a class rating system from 1 to 4.

Factory Acceptance Test and Site Acceptance Test — the two commissioning milestones at which equipment is formally verified before handover.

A network topology used in high-performance data centres that provides full bisection bandwidth by over-provisioning links at higher layers of the hierarchy.

Measure of processor computational throughput, used to compare GPU and HPC system performance.

Using outside ambient air or water temperature to cool a facility without mechanical refrigeration, reducing energy use.

NVIDIA's Blackwell-generation rack-scale AI system — 36 Grace CPUs and 72 Blackwell GPUs interconnected at 130 TB/s via NVLink — delivering 1.4 exaFLOPS of AI performance at roughly 120 kW per rack.

EU Regulation 2016/679 requiring appropriate technical and organisational safeguards for personal data, with direct implications for data residency and datacenter location decisions.

A massively parallel processor originally designed for graphics rendering, now the primary compute engine for AI training and inference at scale.

The share of a data centre energy that comes from green sources, based on onsite generation and the grid renewable ratio.

The portion of grid energy drawn by the data centre attributable to renewable generation in the local grid mix.

The renewable percentage of the local electricity grid: proportion of grid electricity from renewable sources in the relevant market.

A 3D-stacked DRAM standard co-packaged on the GPU die, delivering memory bandwidths of 3–5 TB/s to feed the parallel compute engines of modern AI accelerators.

Rack arrangement separating hot exhaust air from cold supply air to improve cooling efficiency.

Very large data centre (typically 40+ MW) operated by cloud providers at massive scale with highly standardised infrastructure.

Software or firmware that creates and manages virtual machines by abstracting physical hardware resources and arbitrating access between guest operating systems and the underlying CPU, memory, and I/O.

A cooling method in which servers are fully submerged in a dielectric fluid that absorbs heat directly from components.

A high-speed, low-latency network fabric used in AI supercomputer clusters to connect GPUs across nodes, enabling RDMA at speeds up to 400 Gb/s per port.

An out-of-band server management standard that allows remote power control, hardware sensor monitoring, and console access independent of the host operating system via a dedicated Baseboard Management Controller (BMC).

The international standard for Business Continuity Management Systems (BCMS), defining how organisations plan, test, and maintain recovery capability for disruptive incidents.

The international standard for Energy Management Systems, requiring organisations to systematically improve energy performance through policy, objectives and monitoring.

The international standard specifying requirements for an Information Security Management System (ISMS), widely required by enterprise colocation and cloud customers as a condition of contract.

Energy used by IT equipment: servers, storage and networking within the data centre.

The average temperature of air entering IT equipment: a key parameter for cooling setpoint management and ASHRAE class compliance.

An open-source container orchestration platform that automates the deployment, scaling, and management of containerised applications across clusters of servers — increasingly the standard scheduler for AI/ML workloads.

A formal process for assessing, approving and controlling changes to critical infrastructure to prevent unintended outages.

Specialist contractor responsible for integrating all building and IT systems into a single unified management platform.

Mechanical, Electrical and Plumbing — the three engineering disciplines that cover a data centre's physical infrastructure.

An open serial and TCP/IP protocol connecting industrial devices — BMS controllers, CDUs, chillers, generators — enabling register-level read/write of sensor data and control parameters.

Mean Time Between Failures and Mean Time To Repair — the two fundamental reliability metrics for data centre equipment and systems.

An NVIDIA GPU feature that partitions a single physical GPU into up to seven isolated instances, each with dedicated compute, memory and bandwidth resources.

Design model where N systems are needed for full operation and one additional unit provides backup if any N fails.

A file-level storage system connected to a network that allows multiple clients to access shared file systems over NFS or SMB/CIFS — commonly used for shared datasets, home directories, and AI training data staging.

Non-industrial water recovered and reused on site: rainwater harvesting or grey water used in cooling systems.

NVIDIA's proprietary high-bandwidth interconnect for GPU-to-GPU and GPU-to-CPU communication within a server or rack-scale system.

A host controller interface optimised for NAND flash SSDs, providing sub-100 µs latency and over 7 GB/s sequential throughput over PCIe — the performance storage standard for servers, distinct from NVMe-oF which extends the protocol over network fabrics.

A protocol extension that allows NVMe storage commands to be carried over a network fabric (InfiniBand, RoCE or Fibre Channel), providing remote storage with near-local latency.

A flat-namespace storage architecture where data is stored as discrete objects with metadata and a unique key, accessed via HTTP/S APIs (S3-compatible) — used for AI training datasets, backups, and data lakes at petabyte scale.

Device that distributes electrical power from a single high-amperage source to multiple rack-mounted devices.

The ratio of real power (kW) to apparent power (kVA); a measure of how efficiently a load draws current.

A long-term contract between a power generator and a buyer that fixes the price and volume of electricity, typically used to secure renewable energy supply.

The complete path from the grid connection through to the IT equipment, including transformers, switchgear, UPS, generators, busbars and PDUs.

Ratio of total facility power to IT equipment power. PUE 1.0 is perfect; the industry average is ~1.5.

Standard unit of vertical height for rack-mounted equipment. 1U = 44.45 mm (1.75 in).

The design capacity for IT power: the maximum IT load the facility infrastructure is engineered to support.

A network protocol that enables RDMA (remote direct memory access) over standard Ethernet, used in AI clusters as an alternative to InfiniBand.

Maximum acceptable age of data recoverable after a disruption — how much data loss is tolerable.

Maximum acceptable time to restore a service or system after a disruption.

A method of data transfer that allows one computer to directly access the memory of another without involving either machine's CPU or OS, enabling very low latency communication.

Market-based instruments that certify that one megawatt-hour of electricity was generated from a renewable source, used by organisations to support renewable energy claims.

The site own renewable share: total renewable energy divided by DC energy consumption, multiplied by 100.

Renewable energy attributed through Guarantees of Origin certificates or equivalent instruments such as RECs.

Renewable energy generated on the data centre site: typically solar PV, small-scale wind or biogas CHP.

Renewable energy secured through Power Purchase Agreements: contracted volumes of renewable generation attributed to the site.

The GHG Protocol's three categories of greenhouse gas emissions: direct (Scope 1), indirect from purchased energy (Scope 2), and value chain (Scope 3).

Contractual commitment defining minimum acceptable uptime, response times, and service quality a provider must deliver.

The dominant protocol for monitoring and managing networked devices — switches, PDUs, UPS systems — by polling OID values from MIBs or receiving asynchronous trap alerts.

A US auditing standard for service organisations covering security, availability, processing integrity, confidentiality and privacy controls over a defined period.

A two-tier network topology in which every leaf switch connects to every spine switch, providing predictable low-latency east-west traffic paths — the dominant architecture for modern datacentre fabrics.

An electronic switch that transfers a critical load between two independent power sources in under 4 milliseconds.

A dedicated high-speed network providing block-level storage access between servers and storage arrays, appearing to servers as locally attached disks — used for databases, virtualised workloads, and mission-critical applications.

A US telecommunications standard for data centre infrastructure, defining four rated levels (Rated 1–4) for power, cooling, cabling and architecture.

Network switch at the top of a server rack that aggregates all server connections within that rack.

Full lifecycle cost of owning and operating IT infrastructure including capex, opex, cooling, space, and labour.

Drinking-quality water supplied to the site, used primarily in cooling towers, adiabatic coolers and humidification.

The sum of all renewable energy attributed to the data centre: onsite generation, PPAs and Guarantees of Origin.

Total water consumed by the data centre across all sources: potable, non-potable and reused water combined.

Battery-backed power system that maintains continuous supply to critical loads during mains failure.

Four-level system rating data centre design resilience from basic (Tier I) to fault-tolerant (Tier IV).

NVIDIA's GPU architecture after Blackwell, planned for 2026, targeting a further ~4× increase in AI compute per chip and continuing the biennial architecture cadence.

A logical partition of a physical network that segments traffic into independent broadcast domains, used in datacentres to isolate management, storage, production, and customer traffic on shared switching infrastructure.

A software emulation of a physical computer that runs a complete operating system and applications in isolation from other VMs sharing the same physical host, managed by a hypervisor.

A liquid cooling approach that uses elevated water temperatures (30–50°C) at the CDU or cold plate, enabling free cooling for more of the year.

Total waste heat rejected by the data centre: DC energy consumption minus any heat usefully reused.

Waste heat recovered from the data centre and delivered to productive use outside the boundary: district heating, domestic hot water, etc.

Annual site water usage (litres) divided by IT equipment energy (kWh). Lower is better.