GPU Infrastructure Glossary_

Definitions and explanations of key GPU infrastructure terms, grouped by discipline and indexed A–Z.

Browse by category_

GPU Platforms Networking & Fabric Cabling & Fiber Liquid Cooling Power & Distribution Facility & Mechanical Testing & Commissioning

GPU Platforms_

43 terms

GPUs, superchips, baseboards, and the NVIDIA rack-scale systems we deploy.

Backplane

A backplane is a rigid circuit board or interconnect panel that provides shared electrical and data pathways for multiple GPU modules and system components within a single chassis.

Blackwell Ultra (B300)

NVIDIA's Blackwell Ultra GPU, the compute building block of the GB300 NVL72 system.

Blind-Mate Connector

A blind-mate connector is a coupling design that allows two connector halves to engage and lock automatically without requiring direct visual alignment or manual screw fastening.

Blind-Mate Power Connector

A connector system that completes high-current power paths automatically when a module is inserted into its bay without requiring direct manual alignment.

Cable Cartridge

A modular cable-management unit that groups and routes multiple interconnect cables within a dense GPU rack assembly.

Compute node serviceability describes the design and layout features that allow an individual server containing GPUs to be installed, accessed, or removed with minimal disruption to adjacent nodes or rack-level systems.

Compute Tray

A compute tray is a modular chassis unit that houses a set of GPUs, CPUs, memory, and local interconnects within a GPU server rack.

DGX

NVIDIA's integrated GPU server systems designed specifically for AI workloads.

DGX SuperPOD

DGX SuperPOD is a pre-integrated cluster of NVIDIA DGX systems designed for large-scale AI training and HPC workloads.

GB200 NVL72 Rack

The GB200 NVL72 Rack is a factory-integrated rack-scale platform that combines multiple GB200 Grace Blackwell superchips into a single NVLink domain.

GB200 Superchip

The GB200 Superchip is a high-bandwidth, co-packaged GPU+CPU compute module used in NVIDIA's GB200 NVL72 rack-scale systems.

GB300 NVL72

NVIDIA's rack-scale system pairing 72 Blackwell Ultra GPUs with 36 Grace CPUs in one liquid-cooled rack.

GPU Baseboard Management

GPU Baseboard Management refers to the dedicated microcontroller and firmware that monitor and control the GPU module's power, thermal, and error states independently of the host server.

GPU Baseboard Tray

A GPU baseboard tray is a removable mechanical assembly that carries multiple GPUs, their local interconnect fabric, and power delivery components inside a server chassis.

GPU Baseboard Tray Rail

A GPU baseboard tray rail is a structural guide that supports and aligns the GPU baseboard tray during insertion into a server chassis.

GPU Cluster

Networked group of GPU servers working together on parallel computing tasks.

GPU Node Cabling Loom

A GPU node cabling loom is a pre-organized bundle of cables that connects the GPUs, NVSwitch, and host components within a single server node.

GPU Rack

Standardized enclosure housing GPU servers, switches, and power distribution for AI workloads.

GPU Tray

A GPU tray is a modular metal chassis that holds a group of GPUs and their supporting electronics for installation into a GPU server or enclosure.

Grace Blackwell

Grace Blackwell is NVIDIA’s combined CPU+GPU superchip platform, integrating a Grace ARM-based CPU with a Blackwell GPU via a high-speed NVLink-C2C interconnect.

Grace CPU

NVIDIA's ARM-based data center CPU designed to pair with GPUs via NVLink-C2C.

H200 GPU

The H200 GPU is a high-performance accelerator from NVIDIA designed for AI and HPC workloads, succeeding the H100 with increased memory capacity and bandwidth.

HBM Memory

High Bandwidth Memory (HBM) is a 3D-stacked DRAM architecture used in NVIDIA data-center GPUs to provide extremely high memory bandwidth in a compact footprint.

HGX Baseboard

NVIDIA's reference GPU module holding 4 or 8 SXM GPUs connected by NVLink.

HPC

High-Performance Computing using clusters of processors for massive parallel processing.

Liquid-Cooled GPU Node

A liquid-cooled GPU node is a server chassis that uses circulating coolant to remove heat from GPUs and other high-power components, replacing or supplementing traditional air cooling.

MGX Modular Architecture

MGX Modular Architecture is NVIDIA’s rack-scale design that integrates GPU compute, NVLink switching, liquid cooling, and power distribution into a modular frame for data-center deployment.

NVL36

NVL36 is a GPU platform configuration that interconnects 36 NVIDIA GPUs via NVLink within a single rack enclosure.

NVL72

NVIDIA's rack-scale GPU system containing 72 GPUs and 36 Grace CPUs in a single liquid-cooled rack.

NVLink Backplane

A passive copper circuit board assembly that directly interconnects multiple GPUs within a single rack using NVLink lanes, eliminating external cables for GPU-to-GPU traffic.

NVLink Spine

The NVLink Spine is the copper-based backplane or cable harness that directly interconnects all GPUs within a single rack-scale system, such as the NVL72, using NVLink for high-bandwidth GPU-to-GPU communication.

Power Bus Bar (Rack)

A power bus bar is a rigid, conductive metal strip that distributes high-current DC power from a rack’s power distribution unit to multiple GPU servers or switches within the same rack.

Power Interconnect Bar

A conductive bar or rail assembly that distributes DC power from rack-level supplies to multiple GPU nodes or trays.

Rack Bus Bar Alignment

Rack bus bar alignment refers to the positioning of power distribution conductors so that GPU trays or sleds make full electrical contact without mechanical offset.

Rack Manifold Integration

Rack manifold integration is the mechanical and fluid connection of coolant distribution manifolds to GPU cold plates and CDU lines inside a rack during assembly.

Rack Power Shelf

A rack power shelf is a modular power distribution unit that converts high-voltage AC input into multiple low-voltage DC outputs to supply GPU servers and networking equipment within a rack.

Rack Weight Distribution

Rack weight distribution is the placement of component mass within a server rack to keep the assembly stable on its mounting points and within floor loading limits.

Reference Architecture

A reference architecture is a validated blueprint for deploying GPU clusters that specifies hardware, networking, cooling, and software configurations to achieve predictable performance and reliability.

Sled

A sled is a modular, self-contained chassis that holds one or more GPUs and their supporting compute, memory, and cooling components, designed to slide into a larger rack-scale enclosure.

Switch Tray

A switch tray is a modular chassis component that houses the NVSwitch ASICs providing NVLink connectivity between GPUs within a single NVL72 rack.

SXM

NVIDIA's socketed GPU module form factor used on HGX baseboards for high power and NVLink.

SXM Socket

The SXM socket is a high-power, high-bandwidth connector interface used to mount NVIDIA SXM-series GPUs onto a baseboard or system board within a data-center server.

TCO

Total cost of ownership — the full lifetime cost of a GPU cluster beyond hardware purchase.

Networking & Fabric_

60 terms

Interconnects, switches, adapters, and the topologies that tie GPU clusters together.

1.6T Ethernet

1.6T Ethernet is a 1.6 terabit per second aggregate line-rate Ethernet standard for high-bandwidth data-center networking.

400GbE

400 Gigabit Ethernet standard for high-speed data center switch-to-server connections.

800GbE

Next-generation 800 Gigabit Ethernet standard for AI data center fabrics.

Adaptive Routing

Adaptive routing is a dynamic traffic management technique that reroutes data across a network fabric based on real-time congestion or link failure conditions.

Bisection Bandwidth

Bisection bandwidth is the total data capacity across the narrowest point of a network, representing the maximum traffic that can flow between two equal halves of the fabric.

BlueField DPU

NVIDIA's data processing unit that offloads networking, storage, and security from the CPU.

Cable-Length Skew

Cable-length skew is the difference in signal propagation time that arises when cables of unequal physical length are used in a multi-lane or multi-path network link.

Co-Packaged Optics

Co-packaged optics (CPO) integrates optical transceivers directly into the same package as the switch ASIC or GPU module, replacing pluggable front-panel optics.

Compute Fabric

The compute fabric is the high-speed network that connects GPUs within a rack for parallel computation.

Congestion Control

Congestion control is a set of mechanisms that prevent or mitigate network traffic overload by regulating data injection rates when switch buffers approach capacity.

ConnectX

NVIDIA's family of high-speed network adapters for InfiniBand and Ethernet GPU fabrics.

Fabric Cabling Map

A Fabric Cabling Map is a documented diagram that shows the exact routing and port assignments for high-speed interconnect cables forming the GPU cluster network.

Fabric Manager

Fabric Manager is control-plane software that initializes, configures, and monitors the intra-rack GPU interconnect fabric.

Fat-Tree Topology

Network topology providing full bisection bandwidth for GPU cluster switch fabrics.

Fiber Trunk Migration

Fiber trunk migration is the relocation or replacement of pre-terminated fiber optic cable bundles that interconnect racks in a scale-out InfiniBand or Ethernet fabric.

Forward Error Correction

Forward Error Correction (FEC) is a technique that adds redundant data to transmitted signals so the receiver can detect and correct errors without retransmission.

GPUDirect RDMA

Technology letting network adapters read and write GPU memory directly across the fabric.

InfiniBand

High-bandwidth, low-latency network fabric for GPU cluster interconnects.

InfiniBand NDR

InfiniBand NDR is the next-generation high-speed interconnect technology used for GPU-to-GPU scale-out networking in data centers.

InfiniBand XDR

InfiniBand XDR is the latest generation InfiniBand data rate, offering 200 Gbps per lane and doubling the bandwidth of the previous NDR generation.

Leaf Switch

A leaf switch is a top-of-rack or middle-of-row network switch that connects directly to GPU servers and uplinks to spine switches in a leaf-spine fabric topology.

Leaf-Spine Architecture

Two-tier network architecture providing consistent latency and predictable scaling.

Linear Pluggable Optics

Linear Pluggable Optics (LPO) are optical transceivers that remove the digital signal processing (DSP) from the module, relying on the host system's SerDes for signal conditioning.

Link Flap

A link flap is the repeated, rapid transition of a network link between the up and down states, causing instability in the fabric.

Lossless Ethernet

Lossless Ethernet is a networking configuration that prevents packet drops by using priority-based flow control to guarantee reliable data delivery.

Management Network

The management network is a dedicated out-of-band Ethernet infrastructure used for remote monitoring, control, and provisioning of data-center hardware.

NCCL

NVIDIA's library for multi-GPU collective communications in distributed AI training.

Network Interface Card

A Network Interface Card (NIC) is a hardware component that connects a server or GPU node to a data center network via a physical port.

Non-Blocking Fabric

A non-blocking fabric is a network topology where any endpoint can communicate with any other endpoint at full line rate without contention, regardless of traffic patterns.

NVLink

NVIDIA's proprietary high-bandwidth GPU-to-GPU interconnect technology.

NVLink Switch

External switch that extends the NVLink GPU-to-GPU fabric across multiple racks.

NVLink-C2C

High-bandwidth chip-to-chip interconnect between NVIDIA Grace CPUs and GPUs.

NVSwitch

NVIDIA silicon providing all-to-all GPU connectivity via NVLink within a node or rack.

Optical Link Budget Margin

Optical link budget margin is the surplus optical power remaining in a fiber link after subtracting all losses from the transmitter output to ensure the receiver meets its sensitivity target.

Optical Power Budget

The optical power budget is the total allowable loss between a transmitter and receiver in a fiber optic link, ensuring reliable signal detection.

OSFP Transceiver

An OSFP transceiver is a high-speed, octal (8-lane) pluggable optical or copper module used for data-center networking, notably in GPU clusters.

Out-of-Band Management

Out-of-band management is a dedicated network path for administering data center hardware independently of the main data network.

Oversubscription Ratio

The ratio of aggregate downlink port bandwidth to available uplink bandwidth in a switch or fabric topology.

Port-to-Port Latency

Port-to-port latency is the time a packet takes to travel from an ingress port to an egress port on a single network switch or router.

Priority Flow Control

Priority Flow Control (PFC) is a link-level flow control mechanism that pauses traffic on individual priority classes to prevent packet loss in lossless fabrics.

QSFP112 Transceiver

A QSFP112 transceiver is a high-speed pluggable optical or copper module used for data-center networking, supporting up to 112 Gbps per lane.

Quantum-X800

NVIDIA's 800G InfiniBand switch platform for large-scale AI back-end networks.

Rail Switch

A rail switch is a network switch that connects a subset of GPUs within a GPU cluster, typically linking one or more GPU nodes in a single rack to the broader fabric.

Rail-Optimized NIC Placement

Rail-optimized NIC placement positions network interface cards within a server so that each GPU maintains a dedicated path through one rail of the scale-out fabric.

Rail-Optimized Topology

GPU cluster wiring scheme that connects each GPU's NIC to a dedicated leaf-switch rail.

RDMA

Remote Direct Memory Access for GPU-to-GPU communication without CPU involvement.

Retimer

A signal-conditioning chip that regenerates a degraded high-speed serial signal to extend its reach over copper or fiber links.

RoCE

RDMA over Converged Ethernet — low-latency GPU networking on an Ethernet fabric.

SFP Cage Keying

SFP cage keying is the set of mechanical features inside an SFP cage that prevent insertion of an incompatible transceiver module.

Spectrum-X

NVIDIA's Ethernet networking platform tuned for AI workloads as an alternative to InfiniBand.

Spine Switch

A spine switch is a high-radix network switch that forms the core layer of a leaf-spine fabric, connecting to every leaf switch to provide non-blocking, low-latency east-west traffic.

Storage Fabric

A storage fabric is a dedicated high-speed network that connects compute nodes to shared storage systems, separate from the compute or GPU interconnect fabric.

Subnet Manager

The InfiniBand fabric controller that discovers the network topology and programs routing tables for all switches and end nodes.

SuperNIC

A SuperNIC is a high-performance network interface card designed for GPU-to-GPU scale-out communication in AI data centers.

Switch Port Density

Switch port density is the number of network ports available per rack unit (RU) of a top-of-rack, leaf, or spine switch.

Switch Radix

Switch radix is the total number of ports present on a network switch.

Top-of-Rack Switch

A Top-of-Rack (ToR) switch is the network switch installed at the top of each rack to aggregate server traffic and connect to the data center's spine network.

Transceiver Break-In Burn

Transceiver break-in burn is the initial powered-on soak period applied to optical modules after installation to surface early failures before live traffic.

Transceiver Compatibility Matrix

A vendor document that lists which transceiver modules are validated for use with specific switch ports or NICs in a given networking fabric.

Transceiver DDM

Transceiver DDM (Digital Diagnostic Monitoring) is a built-in feature in optical transceivers that reports real-time operating parameters like temperature, voltage, and optical power.

Cabling & Fiber_

69 terms

Fiber, copper, connectors, pathways, and the structured-cabling discipline behind every link.

AEC Cable

Active Electrical Cable bridging the distance gap between DAC and AOC options.

AOC Cable

Active Optical Cable integrating transceivers for medium-distance data center links.

APC vs UPC Polish

APC (Angled Physical Contact) and UPC (Ultra Physical Contact) are two types of fiber optic connector end-face polishes that differ in their return loss characteristics and are used in different signal integrity contexts.

Bend Radius

The tightest curve a fiber or cable can take before signal loss or damage occurs.

Bend-Insensitive Fiber

Bend-insensitive fiber is an optical fiber designed to maintain low signal loss even when bent around tight corners.

BICSI

Professional association setting standards for ICT infrastructure and cabling certification.

Breakout Cable

A breakout cable is a single cable assembly that splits into multiple smaller connectors, enabling a high-density port to connect to several lower-density ports.

Cable Basket Tray

A cable basket tray is an open wire-mesh support used to route and organize copper and fiber cables above or beside racks.

Cable Comb

A cable comb is a slotted tool or fixture that separates and aligns individual cables within a larger bundle.

Cable Labeling Scheme

A cable labeling scheme is a standardized alphanumeric or barcode system applied to both ends of every cable to uniquely identify its source, destination, and circuit path.

Cable Management

Practices and systems for organizing, routing, and maintaining data center cables.

Cable Management Arm

An articulating arm that supports and organizes a server's cables as the chassis slides out of the rack for service.

Cable Pull Tension

Cable pull tension is the force applied to a cable during installation, which must be limited to prevent damage to the conductors, fibers, or jacket.

Cable Slack Loop

A cable slack loop is a controlled service loop of excess cable length left at the rack or tray to allow for future re-termination, rerouting, or equipment movement.

Cable Tray

Overhead or underfloor support system for organized cable pathway routing.

Cabling Pathway Fill

Cabling pathway fill is the occupied portion of a conduit, tray, or raceway relative to its total internal cross-section.

Cassette Polarity Method

The Cassette Polarity Method is a structured cabling technique using pre-terminated fiber cassettes to ensure correct transmit-to-receive pairing across MPO trunk cables.

Connector Cleaning

Connector cleaning is the process of removing contamination from fiber optic or copper connector end-faces to ensure reliable signal transmission.

DAC Cable

Direct Attach Copper cable for short-distance, low-latency data center connections.

Dispersion

Dispersion is the broadening of optical pulses as they travel through fiber, which limits the maximum data rate and distance of a link.

Dust Cap

A dust cap is a protective cover placed over fiber optic connector end faces or ports to block dust and contaminants.

Ferrule

A ferrule is a cylindrical component that precisely aligns and secures an optical fiber within a connector end-face.

Ferrule Pin Alignment

Ferrule pin alignment uses guide pins in multi-fiber connectors to position fiber cores accurately during mating.

Fiber Dark-Fiber Spare

A dark-fiber spare is an installed but unlit optical-fiber strand reserved for future activation or replacement in a structured cabling plant.

Fiber Enclosure

A fiber enclosure is a protective housing that organizes, stores, and manages fiber optic cable connections and splices within a data center rack or bay.

Fiber End-Face Inspection

Fiber end-face inspection is the process of examining the polished tip of a fiber-optic connector under magnification to verify it is clean and free of defects before mating.

Fiber Jumper

A fiber jumper is a short, pre-terminated optical cable with connectors at both ends used to link network ports or patch panels.

Fiber Patch Panel

Rack-mounted enclosure providing organized termination for fiber optic cables.

Fiber Polarity

Fiber polarity ensures that the transmit and receive fibers are correctly aligned between transceivers at each end of a link.

Fiber Raceway

A fiber raceway is a physical pathway or trough system used to route, protect, and organize fiber optic cables within a data center rack or row.

Fill Ratio

Fill ratio is the percentage of a cable tray or conduit's cross-sectional area occupied by cables, ensuring proper airflow and thermal management.

Fusion Splice

A fusion splice is a permanent, low-loss joint between two optical fibers created by melting and fusing their ends together using an electric arc.

Innerduct

Innerduct is a flexible sub-duct installed inside a larger conduit to organize and protect fiber optic cables.

Ladder Rack

An open, ladder-style overhead cable pathway used to support and route bundles of data and power cables in data-center aisles.

LC Connector

A small-form-factor fiber optic connector using a 1.25 mm ferrule, widely used in structured cabling for GPU data-center networks.

Loopback Plug

A loopback plug is a passive connector that routes a port's transmit signal directly back to its receive path for isolated testing.

Male vs Female MPO

Male and female MPO connectors are distinguished by alignment features: the male version carries two guide pins while the female version contains the corresponding holes.

Mechanical Splice

A permanent or semi-permanent fiber-optic connection made by aligning and joining two bare fiber ends without a connector.

Modal Bandwidth

Modal bandwidth is the product of the bandwidth and distance over which a multimode fiber can carry a signal before dispersion degrades it.

MPO Cassette

An MPO cassette is a fiber-optic adapter module that breaks out a multi-fiber MPO connector into individual LC or SC duplex connectors for easier patching.

MPO Connector

High-density fiber connector terminating 8–24 fiber strands in a single connector.

MPO-16 Connector

An MPO-16 connector is a multi-fiber push-on optical connector that aligns and terminates sixteen fibers within a single rectangular ferrule and housing.

MPO-24 Connector

An MPO-24 connector is a multi-fiber push-on optical connector that terminates twenty-four fibers within a single rectangular ferrule.

MTP Connector

An MTP connector is a high-density, multi-fiber push-pull connector used for parallel optical interconnects in data center cabling.

Multimode Fiber

A type of optical fiber with a larger core diameter (typically 50 µm) used for short-reach, high-bandwidth links within a data center, commonly OM3, OM4, or OM5 grades.

OM3 Fiber

OM3 is a multimode optical fiber optimized for 850 nm VCSEL-based transmission, commonly used for short-reach data center links.

OM4 Fiber

Laser-optimized multimode fiber supporting 400 Gb/s for intra-data-center connections.

OM5 Fiber

Wideband multimode fiber supporting SWDM for higher bandwidth density.

One-Jumper Reference Method

The one-jumper reference method sets the zero-loss baseline for optical power measurements by placing a single reference jumper between the light source and power meter.

Optical Attenuation

Optical attenuation is the reduction in signal power as light travels through a fiber optic cable, connectors, or splices.

Optical Transceiver

A pluggable module that converts electrical signals into optical signals for transmission over fiber optic cables in data center networks.

OS2 Fiber

Single-mode fiber for long-distance backbone connections in data centers.

OSFP

Transceiver form factor with better thermal management for 400G/800G+ connections.

Overhead Fiber Runner

An overhead fiber runner is an elevated cable pathway installed above racks to route and support fiber optic cables in data centers.

Parallel Optics

Parallel optics transmit data simultaneously over multiple fiber strands in a single ribbon cable, typically terminated with MPO connectors, enabling high-bandwidth short-reach links.

Patch Cord

A patch cord is a short, flexible cable with connectors on both ends used to connect devices within a rack or patch panel.

Pigtail

A pigtail is a short, factory-terminated fiber optic cable with a connector on one end and bare fiber on the other, used for splicing to field cables.

Pre-Terminated Trunk

A pre-terminated trunk is a factory-assembled, multi-fiber cable assembly with connectors already installed on both ends, ready for direct installation.

QSFP-DD

Transceiver form factor supporting 400G/800G Ethernet, backwards compatible with QSFP.

Reference Test Cord

A reference test cord is a high-quality fiber optic patch cord employed to establish a zero-loss baseline before measuring insertion loss or return loss on installed cabling.

Single-Mode Fiber

Single-mode fiber is an optical fiber with a small core that transmits a single light mode, enabling high-bandwidth, long-distance links in data center networks.

Strain Relief

A mechanical feature that anchors a cable near its connector to prevent tension from reaching the solder or splice points.

Structured Cabling

Standardized cable systems providing physical network infrastructure for data centers.

Structured Cabling MACs

Structured cabling MACs are the moves, adds, and changes performed on the permanent cabling infrastructure after initial installation.

Structured Cabling Trunk

A factory-terminated multi-fiber MPO trunk cable that runs between patch panels to enable rapid, high-density fiber deployment in data centers.

TIA-606 Labeling Standard

ANSI/TIA standard for labeling and documenting telecom and data-center cabling infrastructure.

TIA-942

Telecommunications Industry Association standard for data center infrastructure design.

Twinax Cable

A twinaxial copper cable used for short-range, high-speed data transmission within a rack, commonly for NVLink GPU-to-GPU connections.

Velcro Cable Wrap

A reusable fabric strap with a hook-and-loop closure used to bundle and manage cables in data center racks.

Liquid Cooling_

51 terms

Cold plates, manifolds, CDUs, and the liquid loops that cool high-density GPU racks.

Approach Temperature

Approach temperature is the temperature difference between the coolant entering a cold plate and the component it cools, indicating heat exchanger efficiency.

Biological Growth Control

Biological growth control manages microbial contamination in liquid cooling loops to prevent biofilm formation and flow restriction.

Cold Plate

Metal heat exchanger mounted directly on GPU/CPU dies for liquid cooling.

Cold Plate Fouling

Cold plate fouling is the gradual buildup of mineral scale, particulates, or corrosion products on the internal fluid channels and heat-transfer surfaces of a liquid-cooling cold plate.

Cold Plate Microchannel

A cold plate with microscopic internal channels that circulates liquid coolant directly beneath the GPU die to remove heat at high density.

Coolant Bypass Loop

A secondary flow path in a liquid-cooling circuit that routes coolant around selected cold plates or servers while the remainder of the system stays online.

Coolant Conductivity

Coolant conductivity measures the ability of the liquid cooling fluid to conduct electrical current, indicating its purity and risk of galvanic corrosion.

Coolant Corrosion Inhibitor

A chemical additive introduced into liquid coolant to limit corrosion of metallic surfaces exposed to the fluid.

Coolant Distribution Manifold

A Coolant Distribution Manifold (CDM) is a liquid-cooling component that distributes coolant to and from multiple cold plates or heat exchangers within a rack.

Coolant Distribution Unit (CDU)

Heat exchanger between rack-level cooling loops and facility water systems.

Coolant Distribution Unit Sidecar

A Coolant Distribution Unit Sidecar is an auxiliary module that augments the fluid handling capacity of a primary CDU in liquid-cooled rack systems.

Coolant Fill and Bleed

Coolant Fill and Bleed is the process of introducing liquid coolant into a data-center liquid cooling loop and removing trapped air to ensure proper circulation and heat transfer.

Coolant Filtration

Coolant filtration removes particulates and contaminants from the liquid cooling loop to prevent blockages and maintain thermal performance.

Coolant Flow Rate

Coolant flow rate is the volume of liquid coolant circulated per unit time through a liquid cooling system to remove heat from GPUs and other components.

Coolant Hose Routing

Coolant hose routing is the planned path and physical arrangement of flexible tubes that carry liquid coolant between the CDU and the cold plates inside a GPU server rack.

Coolant Inventory

Coolant Inventory is the total volume and type of dielectric or water-based fluid charged into a liquid-cooled GPU cluster’s closed-loop system.

Coolant Loop Pressure Test

A coolant loop pressure test verifies the liquid cooling system's ability to hold a specified pressure without leaks before the rack is commissioned.

Coolant Pump Redundancy

Coolant pump redundancy maintains coolant circulation in a liquid cooling system when one pump stops operating.

Coolant Return Temperature

Coolant return temperature is the measured temperature of liquid after it has absorbed heat from GPUs or other components and is flowing back toward the cooling distribution equipment.

Coolant Supply Temperature

Coolant supply temperature is the temperature of the liquid coolant as it enters the cold plates or heat exchangers in a liquid-cooled GPU rack.

Cooling Capacity Margin

The difference between installed liquid-cooling capacity and the actual thermal load presented by the IT equipment.

Cooling Loop Flushing

Cooling loop flushing circulates a cleaning fluid through liquid-cooling circuits to remove manufacturing residues, particulates, and oxides before introducing the final coolant.

Cooling Manifold

In-rack distribution header that splits coolant from the CDU to each server tray and returns it.

Deionized Water Coolant

Deionized water coolant is purified water with dissolved ions removed for circulation in direct-to-chip liquid cooling loops.

Dew Point Control

Dew point control maintains liquid coolant temperature above the dew point of ambient air to prevent condensation on cooled surfaces.

Differential Pressure

Differential pressure is the measured difference in fluid pressure between two points in a liquid cooling loop, indicating flow resistance and system health.

Direct Liquid Cooling (DLC)

Thermal management using liquid coolant through cold plates on GPUs and CPUs.

Direct-to-Chip Cooling

Liquid cooling that delivers coolant to cold plates mounted on each GPU and CPU die.

Drip Tray

A drip tray is a containment component installed beneath liquid-cooled computing equipment to capture and route potential coolant leaks away from sensitive electronics.

Dripless Quick Disconnect

A dripless quick disconnect is a coupling that automatically seals both halves when separated, preventing coolant spillage during maintenance.

Facility Supply/Return Headers

Facility supply and return headers are the primary distribution pipes that carry chilled coolant from central plant equipment to racks and return warmed coolant for reconditioning.

Facility Water Loop

The building-side cooling water circuit that rejects heat collected by the CDU.

Heat Rejection

Heat rejection is the process of transferring waste thermal energy from the liquid cooling loop to the ambient air or facility water loop.

Hybrid Cooling

Hybrid cooling combines direct-to-chip liquid cooling with traditional air cooling to manage GPU heat loads in a single rack.

Immersion Cooling

Cooling method that submerges servers in thermally conductive dielectric fluid.

In-Rack CDU

An in-rack CDU is a coolant distribution unit mounted inside a rack that circulates conditioned liquid between facility supply lines and the server cooling loops.

Leak Detection

Leak detection refers to the system of sensors and monitoring equipment that identifies coolant leaks in liquid-cooled racks and coolant distribution units (CDUs).

Liquid-to-Air CDU

A Liquid-to-Air CDU (Coolant Distribution Unit) is a heat exchanger that transfers heat from the coolant loop to the data center room air, typically used in rear-door heat exchanger or direct-to-chip cooling loops.

Liquid-to-Liquid CDU

A liquid-to-liquid CDU is a heat exchanger that transfers heat from a facility’s coolant loop to a secondary coolant loop serving the rack’s liquid-cooled components.

Manifold Blanking Plug

A manifold blanking plug is a sealing component inserted into an unused liquid cooling manifold port to prevent coolant leakage and maintain system integrity.

Manifold Flow Balancing

Manifold flow balancing equalizes coolant volume across parallel branches feeding multiple cold plates in a liquid-cooled rack.

Primary Fluid Network

The primary fluid network is the main coolant loop that carries heat from the rack’s liquid-cooled components to the facility’s heat rejection system.

Propylene Glycol Coolant

A heat-transfer fluid used in liquid-cooled data centers to lower the freezing point and inhibit corrosion in the coolant loop.

Quick Disconnect (QD)

Dripless coupling that connects liquid-cooling lines to a server tray without spilling coolant.

Rear-Door Heat Exchanger (RDHx)

Liquid-cooled door on the back of a rack that removes heat from exhaust air before it enters the room.

Row-Based CDU

A row-based CDU is a coolant distribution unit that supplies secondary-loop liquid to multiple racks arranged in a single row.

Secondary Fluid Network

The secondary fluid network is the liquid cooling loop that carries heat from the rack’s coolant distribution unit (CDU) to the cold plates or heat exchangers inside the servers.

Technology Cooling System (TCS)

The closed, treated server-side liquid loop that circulates coolant through cold plates, isolated from facility water by the coolant distribution unit (CDU).

Thermal Interface Material (TIM)

Material filling air gaps between chip dies and heat sinks or cold plates.

Two-Phase Cooling

Cooling that absorbs heat by boiling a coolant from liquid to vapor at the heat source.

Vertical Cooling Manifold

A vertical cooling manifold is a rack-mounted pipe assembly that supplies coolant to and collects warmed fluid from multiple liquid-cooled nodes stacked vertically.

Power & Distribution_

44 terms

PDUs, busway, power whips, and the bonding that feeds dense GPU rows safely.

2N Redundancy

2N redundancy is a fault-tolerant power architecture where each critical load is served by two independent, fully-rated power paths, so that any single component or feed failure leaves the other path to carry the full load without interruption.

415V Distribution

415V distribution is a three-phase electrical power system used in data centers to deliver higher voltage to rack-level equipment, reducing current and copper losses.

Automatic Transfer Switch

An Automatic Transfer Switch (ATS) automatically switches a load between primary and backup power sources without manual intervention.

Branch Circuit Monitoring

Branch Circuit Monitoring (BCM) continuously measures electrical parameters on individual power circuits within a rack or row to track load, voltage, and current.

Busbar

A busbar is a rigid metallic conductor that distributes high-current power within a rack or row of data-center equipment.

Busway

Prefabricated power distribution system using enclosed bus bars along data center rows.

Busway Tap-Off Box

A pluggable electrical enclosure with an integrated circuit breaker that connects a rack PDU to an overhead busway for power distribution.

Circuit Breaker Coordination

Circuit breaker coordination is the selective arrangement of breakers so that only the faulted circuit is isolated, leaving upstream feeders live.

Equipment Grounding Conductor

A conductor that connects non-current-carrying metal parts of equipment to the system grounded conductor or grounding electrode, ensuring a low-impedance path for fault current.

Feeder Cable Sizing

Feeder cable sizing determines the conductor cross-section and insulation rating needed to carry the calculated load current from the upstream distribution board to rack-level PDUs without exceeding thermal or voltage limits.

Ground Fault Detection

Ground fault detection identifies unintended current flow from conductors to ground in AC power distribution circuits.

Grounding & Bonding

The electrical safety and signal-reference system tying equipment to a common low-impedance ground.

Grounding Busbar (TGB/TMGB)

A copper or aluminum bar in the data center that provides a common bonding point for racks, cable trays, and equipment to ensure a low-impedance path to earth ground.

Grounding Electrode System

A grounding electrode system is the set of electrodes and interconnecting conductors that establish a direct low-impedance connection between the facility electrical system and earth.

Harmonic Distortion

Harmonic distortion is the presence of unwanted frequency components that alter the shape of an AC voltage or current waveform away from a pure sinusoid.

Inrush Current

Inrush current is the brief, high surge of electrical current drawn by power supplies and equipment when first energized.

IT PDU vs Facility PDU

An IT PDU supplies power to individual servers and accelerators inside a rack while a facility PDU handles upstream distribution from building electrical sources to multiple racks or rows.

Metered Rack PDU

A metered rack PDU is a rack-mounted power distribution unit that measures and reports aggregate electrical consumption at the rack level.

N+1 Redundancy

N+1 redundancy is a power architecture where a system includes one extra component beyond the minimum needed to support the load, ensuring continued operation if any single component fails.

Neutral Conductor Sizing

Neutral conductor sizing determines the cross-section of the return conductor in three-phase AC distribution so it can carry unbalanced and harmonic currents without overheating.

Overhead Busway

Overhead busway is a suspended modular busbar system that routes electrical power above racks to feed rack PDUs.

Phase Balancing

Phase balancing distributes electrical load evenly across all three phases of a three-phase power system to prevent overload and improve efficiency.

Power Capping

Power capping is a feature that limits the maximum electrical power a GPU or server can draw, preventing overloads and managing data-center power budgets.

Power Density

Power density is the amount of electrical power consumed per unit of rack floor area or volume, typically expressed in kilowatts per rack or per square foot.

Power Distribution Unit (PDU)

Device that distributes electrical power from facility circuits to rack equipment.

Power Feed A/B Diversity

Power feed A/B diversity supplies two electrically independent paths from separate upstream sources to the same rack or chassis.

Power Monitoring CT

A current transformer that steps down high line currents to safe measurable levels for power monitoring equipment.

Power Usage Effectiveness Budget

Power Usage Effectiveness (PUE) Budget is the target ratio of total facility power to IT equipment power that a data center must maintain to meet operational and efficiency goals.

Power Whip

Pre-assembled electrical cable connecting rack PDUs to facility power sources.

Power Whip Connector

A power whip connector is the interface that attaches a flexible, pre-terminated cable assembly from a rack’s power distribution unit to a GPU server or switch power supply.

Rack PDU Phase Rotation

Rack PDU phase rotation is the procedure of verifying and aligning the three-phase power feed sequence to a rack’s power distribution unit to ensure correct phase orientation before energizing equipment.

Rack Power Provisioning

Rack power provisioning supplies electrical power from facility feeds to all equipment inside a data-center rack.

Rack Power Whip Rating

The rack power whip rating specifies the maximum continuous current and voltage capacity of the flexible cable assembly that feeds facility power into a rack PDU.

Redundant Power Feed

A redundant power feed supplies a rack or component with two independent electrical paths so that a single feed failure does not cause an outage.

Remote Power Panel

A Remote Power Panel (RPP) distributes three-phase AC power from the facility’s main feed to individual rack PDUs in a data center row.

Selective Coordination

Selective coordination arranges overcurrent protective devices so that a fault trips only the nearest upstream device.

Static Transfer Switch

A Static Transfer Switch (STS) automatically switches a load between two independent power sources without mechanical contact, ensuring near-instantaneous failover.

Supplementary Bonding

Supplementary bonding is an intentional electrical connection between exposed conductive parts of equipment and extraneous conductive parts to prevent dangerous voltage differences in a data center.

Switched Rack PDU

A switched rack PDU is a rack-mounted power distribution unit that provides remote on/off control of individual outlets.

Switchgear

Switchgear is an assembly of electrical disconnect switches, fuses, and circuit breakers used to control, protect, and isolate electrical equipment in a data center’s power distribution system.

Three-Phase Power

Three-phase power is an electrical distribution method using three alternating currents that overlap to deliver a more consistent and higher-capacity power supply than single-phase.

Uninterruptible Power Supply

An Uninterruptible Power Supply (UPS) provides emergency backup power and conditions incoming electrical supply to protect critical data-center equipment from outages, sags, and surges.

Voltage Drop

Voltage drop is the reduction in electrical potential along a conductor due to its resistance and the current flowing through it.

Whip Drop

A whip drop is a short, flexible power cable assembly that connects from an overhead busway or power distribution unit to a single rack.

Facility & Mechanical_

45 terms

Containment, air handling, heat rejection, and the building systems around the white space.

Air Containment Curtain

A flexible barrier that separates hot and cold air paths within a data center aisle to improve cooling efficiency.

Air Economization Hours

Air economization hours quantify the annual periods when outside air can be used for data-center cooling instead of mechanical refrigeration.

Air Handling Unit

An Air Handling Unit (AHU) is a mechanical device that conditions and circulates air as part of a data center's HVAC system.

Aisle Pressure Differential

Aisle pressure differential is the static air pressure difference maintained between a contained cold aisle and the adjacent hot aisle or room space.

Blanking Panel

A blanking panel is a solid or vented plate installed in an empty rack unit to maintain proper airflow and prevent recirculation of hot exhaust air into the cold aisle.

Cabinet Grounding Kit

A collection of conductors, busbars, and hardware that bonds a server cabinet frame to the facility grounding system.

Cable Penetration Seal

A cable penetration seal is a firestop device that closes gaps around cables passing through walls, floors, or ceilings to maintain fire ratings and prevent smoke spread.

Chilled Water Plant

A chilled water plant is a centralized mechanical system that produces and distributes chilled water to cool data center equipment, including GPU clusters.

Cold Aisle Containment

Cooling strategy that encloses the cold supply side of server racks.

Computer Room Air Conditioner

A Computer Room Air Conditioner (CRAC) is a dedicated cooling unit that maintains temperature and humidity within a data center by circulating chilled air through a raised-floor plenum or overhead ductwork.

Containment Bypass Airflow

Containment bypass airflow is unintended air movement that circumvents hot-aisle or cold-aisle separation barriers in a data center.

Cooling Tower

An evaporative heat-rejection device that removes heat from the facility water loop in liquid-cooled GPU data centers.

CRAH

Computer Room Air Handler — a chilled-water unit that cools and circulates data center air.

DCIM

Data Center Infrastructure Management (DCIM) is a software and hardware platform that monitors, measures, and manages the power, cooling, and physical assets of a data center in real time.

Delta T

Delta T is the temperature difference between the coolant entering and leaving a liquid-cooled component or system.

Dry Cooler

A finned-coil heat exchanger that rejects heat from the facility liquid loop to ambient air without evaporating water.

Economizer

An economizer is a mechanical system that uses outside air to cool a data center when ambient conditions are favorable, reducing reliance on mechanical refrigeration.

Environmental Monitoring

Environmental monitoring is the continuous measurement of temperature, humidity, and airflow within the data center to ensure operating conditions stay within equipment specifications.

Fire Suppression System

A fire suppression system detects and extinguishes fires in data center spaces while minimizing damage to electronic equipment.

Floor Loading

The structural weight capacity of a data center floor, verified before placing heavy GPU racks.

Floor Loading Survey

A floor loading survey assesses the structural capacity of a data center floor to support the static and dynamic weight of deployed hardware.

Free Cooling

Free cooling uses outside air or other natural heat sinks to cool data center equipment, reducing or eliminating mechanical refrigeration.

Gray Space

Gray space is the intermediate zone between the white space (data floor) and the black space (utility infrastructure) where facility mechanical and electrical systems are distributed to support the IT load.

Hot Aisle Containment

Cooling strategy that encloses server rack exhaust to improve cooling efficiency.

Hot Aisle Plenum Return

A hot aisle plenum return is an overhead or raised-floor air pathway that collects and exhausts hot exhaust air from the rear of GPU servers back to the cooling system.

In-Row Cooler

An air-handling unit installed between server racks within a row to provide close-coupled cooling for high-density GPU clusters.

Loading Dock Staging

Loading dock staging is the temporary holding and preparation area where incoming GPU system crates and racks are received, inspected, and organized before being moved to the data center floor.

Mechanical Electrical Plumbing

Mechanical Electrical Plumbing (MEP) refers to the building systems that provide power, cooling, and water to data center racks.

Overhead Structural Support

Overhead structural support refers to ceiling-suspended frameworks that carry the weight of data-center cabling, cooling distribution, and ancillary equipment above the rack rows.

Plenum

Plenum refers to the air-handling space above a drop ceiling or below a raised floor used for HVAC return or supply air, which in data centers must meet fire and smoke safety codes for cabling and materials.

PUE

Power Usage Effectiveness: ratio measuring data center energy efficiency.

Rack Anchoring

Rack anchoring is the process of securing server racks to the data-center floor or overhead structure to prevent movement during seismic events or normal operations.

Rack Ingress Protection

Rack ingress protection describes the enclosure features and ratings that limit entry of dust, liquids, and foreign objects into IT racks.

Rack Levelling

Rack levelling is the process of adjusting a server rack’s feet to achieve a plumb and level frame on the data-center floor.

Rack Row Alignment

Rack row alignment is the arrangement of equipment racks into straight, parallel lines with uniform spacing and orientation across a data hall floor.

Rack Unit (U)

Standard unit of measure for vertical space in a server rack (1.75 inches).

Raised Floor

Elevated floor system creating a plenum for air distribution and cable routing.

Raised Floor Grommet

A raised floor grommet is a protective insert that seals the cable cutout in a raised floor tile, allowing cables to pass through while maintaining airflow and debris containment.

Seismic Bracing

Seismic bracing is structural hardware that secures racks, cable trays, and piping to the building to prevent overturning or shifting during an earthquake.

Structural Floor Load Rating

Structural floor load rating is the maximum distributed weight a data center floor can support without risk of deflection or failure.

Supply Air Temperature

Supply air temperature is the temperature of conditioned air delivered by a cooling unit to the inlet face of IT equipment.

Under-Floor Pressure

Under-floor pressure is the static air pressure maintained in the raised-floor plenum that drives cooling supply air through perforated tiles into equipment intakes.

Vapor Barrier

A vapor barrier is a continuous layer of low-permeance material that limits water-vapor diffusion through walls, floors, or enclosures.

Water Detection Rope

Water detection rope is a linear sensor cable that identifies liquid contact along its length by changes in electrical properties between conductors.

White Space

White space is the area within a data center that is specifically allocated for IT equipment, such as GPU servers and networking gear, as opposed to support infrastructure.

Testing & Commissioning_

48 terms

POST, burn-in, certification, and the verification that turns a powered-on cluster into an accepted one.

Acceptance Testing

Acceptance Testing is the formal process of verifying that a deployed GPU system meets specified performance and reliability criteria before handover to operations.

As-Built Documentation

As-Built Documentation is the final set of drawings, diagrams, and records that capture the actual physical configuration and connections of a deployed system after installation and commissioning.

Bit Error Rate Test

A Bit Error Rate Test (BERT) measures the ratio of erroneous bits received to total bits transmitted over a digital link to assess signal integrity.

BMC (Baseboard Management Controller)

An out-of-band management processor on a server that provides remote power control, health monitoring, and console access independent of the main operating system.

Burn-In Testing

Sustained high-load testing after deployment to identify early hardware failures.

Cable Certification Report

A Cable Certification Report is a formal document verifying that installed cabling meets specified performance standards.

Cable Certification Tester

A test instrument that verifies installed copper and fiber cabling links against transmission performance requirements.

Cluster Validation

Cluster validation is the systematic testing process that confirms a GPU cluster meets performance, connectivity, and stability requirements before production use.

Commissioning (Cx)

The structured process of verifying that an installed GPU cluster works and meets its design before production.

Commissioning Levels

Commissioning levels are staged verification phases that confirm installation quality and operational readiness of GPU racks and clusters before handover.

Continuity Test

A continuity test verifies that an electrical or optical path is complete and unbroken from end to end.

Coolant Leak Test Sign-Off

A documented approval confirming that liquid cooling circuits have passed leak detection before operational commissioning.

Coolant Quality Verification

Coolant Quality Verification is the process of testing dielectric fluid or water-glycol mixture properties to ensure they meet OEM specifications before energizing liquid-cooled GPU systems.

ECC Memory

Error-correcting memory that detects and fixes bit errors, critical for long GPU training runs.

Encircled Flux Compliance

Encircled flux compliance defines a controlled launch condition for multimode fiber testing that produces repeatable attenuation and bandwidth results.

Fabric Link Sweep

A commissioning procedure that systematically checks connectivity and basic integrity across every link in a GPU cluster's scale-out fabric network.

Factory Acceptance Test

A Factory Acceptance Test (FAT) is a formal verification performed at the manufacturer's facility to confirm that a system meets its design specifications before shipment.

Fiber Certification

Formal testing and documentation verifying fiber installations meet performance standards.

Fiber Inspection Scope

A handheld microscope used to examine the end faces of fiber optic connectors for contamination, scratches, or defects before mating.

Firmware Baseline

The agreed set of firmware versions for BIOS, BMC, NICs, and GPUs that a cluster is updated to before acceptance testing begins.

Functional Performance Test

A Functional Performance Test verifies that a GPU data-center system meets its specified operational and throughput benchmarks under load.

GPU Stress Validation

GPU stress validation applies sustained high-utilization workloads to newly installed GPUs to confirm stable operation under load.

IEC 61300-3-35 (Fiber Endface Inspection)

International standard defining pass/fail criteria for the cleanliness of optical fiber endfaces.

Insertion Loss

Measurement of optical signal power lost through a fiber connection, in decibels.

Integrated Systems Test

An Integrated Systems Test validates that all subsystems within a GPU data-center rack (power, cooling, networking, compute) operate correctly together as a single functional unit.

Level 5 Commissioning

Level 5 Commissioning is the integrated system-level testing of the fully assembled rack, including power, cooling, and networking, to verify it operates as a cohesive unit before handover.

Link Integrity Test

A Link Integrity Test verifies that every physical data link in the GPU cluster—copper NVLink traces and fiber optic scale-out cables—is electrically or optically continuous and correctly terminated.

Link Loss Budget Calculation

Link loss budget calculation sums the expected attenuation sources in an optical channel to confirm that total loss stays within the transceiver's operating range.

Load Bank Testing

Load bank testing is a commissioning procedure that applies an artificial electrical load to a power system to verify its performance under stress.

Megger Test

A Megger test measures insulation resistance of electrical wiring and components using a high-voltage megohmmeter.

Method of Procedure

A method of procedure is a written, step-by-step document that defines the approved sequence for executing a specific task during installation, testing, or commissioning.

NCCL Bandwidth Test

A diagnostic tool that measures the aggregate data-transfer rate between GPUs using NVIDIA’s Collective Communications Library.

Optical Loss Test Set

An optical loss test set (OLTS) measures the end-to-end insertion loss of a fiber optic link to verify it meets the required performance budget.

OTDR Testing

Fiber optic testing that maps cable characteristics using reflected light pulses.

Polarity Verification

Polarity verification is the process of confirming that optical fiber connectors are aligned correctly to ensure proper transmit-to-receive signal paths.

POST Verification

Power-On Self-Test confirming correct hardware initialization after deployment.

Power-On Self-Test Sequence

The Power-On Self-Test (POST) sequence is the initial diagnostic process a GPU server runs at power-up to verify hardware integrity before loading the operating system.

Pre-Power Inspection

A systematic verification of rack assembly, connections, and safety conditions performed before applying electrical power to GPU systems.

Pressure Decay Test

A non-destructive leak test that pressurizes a liquid cooling loop and monitors for pressure drop to detect leaks before system energization.

Punch List

A punch list is a documented checklist of incomplete or defective items that must be resolved before a system is formally accepted.

PXE Boot

A network-based boot method used to provision operating system images to GPU nodes during initial bring-up and commissioning.

Redfish

Redfish is the DMTF standard RESTful API for out-of-band management of servers and data center infrastructure.

Redline Drawings

Redline drawings are the design documents marked up on site to record all deviations from the issued drawings that occurred during installation.

Return Loss

Measurement of light reflected back at a fiber connection, in decibels.

Site Acceptance Test

A Site Acceptance Test (SAT) is a formal procedure performed at the deployment site to verify that installed hardware and systems meet contractual specifications before final acceptance.

Thermal Imaging Survey

A non-contact inspection technique using infrared cameras to detect temperature anomalies in electrical and cooling components during commissioning.

Time-Domain Reflectometry

Time-domain reflectometry sends a pulse along a conductor and measures the timing and amplitude of reflections to identify impedance changes or discontinuities.

Torque Verification

Torque verification confirms that mechanical fasteners meet the tightening values given in the OEM documentation.

Browse A–Z_

1 2 4 8 A B C D E F G H I L M N O P Q R S T U V W

1