AI Glossary / Data engineering and architecture / Intermediate

Fabric computing

Definition

Fabric computing is an infrastructure approach that connects computing resources through a shared interconnection fabric and manages them as a coordinated system. Depending on the design, the resources can include servers, storage access, networking and other modular components.

What it means

The fabric is the connectivity structure through which resources communicate. It is commonly built from switches, links and interfaces. Management software can configure relationships and policies across the connected resources, reducing the need to treat each connection as an isolated setup task.

The terminology varies. Gartner's narrower fabric-based computing definition emphasizes modular components designed around a fabric, while vendor descriptions of fabric computing often emphasize integrated compute, networking and storage access. A specific product's architecture must therefore be examined rather than inferred from its label.

Fabric computing is also different from a data fabric, which concerns integrated data management and access. A computing fabric provides infrastructure connectivity; it does not automatically reconcile data definitions or create a business knowledge layer.

How it works

Resources attach to the fabric through compatible interfaces. The topology and switching mechanisms determine available paths, bandwidth and latency. A management layer discovers resources and applies supported configuration policies, such as network connectivity or server profiles.

The extent of pooling or composition depends on the implementation. Some systems coordinate complete servers and their input/output connections. Others disaggregate particular resources so that they can be assigned more flexibly. The word fabric does not guarantee that any processor can directly use any memory or accelerator as though it were locally attached.

Shared connectivity introduces shared capacity and failure considerations. Several workloads can compete for the same links, and a failed component can affect multiple paths. Redundant designs can improve resilience, but their actual behavior depends on topology, protocols and configuration.

Applications still need appropriate software and data placement. Connecting resources efficiently does not make a serial program parallel or eliminate the cost of transferring large datasets.

Three practical examples

• A data-center team applies supported server and connectivity profiles when assigning compatible computing resources to a new workload.
• An analytics cluster uses a switched interconnect to exchange data among nodes and access shared storage services.
• An AI infrastructure team evaluates how accelerator servers, storage paths and network links work together during model loading and distributed computation.

Think of it this way

Think of a well-connected industrial campus whose buildings share a planned transport network and coordinated access arrangements. The network makes movement and reassignment easier, but each road still has a capacity and each building has a purpose.

The analogy is limited because computing fabrics use precise communication protocols and can operate at very different scales. Shared connectivity does not turn every resource into one physically unlimited machine.

Worked example

Checking the shared link before adding servers

An invented analytics setup has four servers, each capable of requesting data at four gigabytes per second. All four share an upstream path that can deliver ten gigabytes per second in aggregate, before additional overheads are considered.

Their combined requested rate is 4 × 4 = 16 gigabytes per second. The shared path cannot satisfy all requests at that rate. If capacity is divided equally under these simplified assumptions, each receives at most 10 ÷ 4 = 2.5 gigabytes per second.

Adding two more servers does not enlarge the same upstream path. Equal sharing among six would yield about 1.67 gigabytes per second per server. Real allocation can differ with traffic patterns and policies, so these figures are bounds for the illustration rather than measured performance.

The example shows why a fabric design must be evaluated end to end. Counting server ports or adding computing capacity alone can conceal a shared bottleneck.

Illustrative case study

A fictional research-services company plans a shared environment for data preparation and model evaluation. Its infrastructure team initially compares server processor counts. A workload review shows that several jobs repeatedly load large datasets through the same storage-facing path.

The team includes that path in its capacity model and tests representative overlapping jobs. It documents which failures the redundant connections can tolerate and what happens when traffic moves to the remaining path. A healthy failover can still reduce available throughput.

The company also distinguishes management convenience from application performance. Applying a configuration profile faster does not by itself shorten a model evaluation. The case illustrates how fabric connectivity, resource management and workload behavior must be considered together; it does not endorse a particular platform.

Common misconceptions and mistakes

• Fabric computing means data fabric. The terms concern different layers and can exist independently.
• Connected resources have unlimited shared capacity. Links and endpoints retain finite throughput and latency.
• Every fabric provides universal memory pooling. Resource composition depends on the architecture.
• Redundant links guarantee unchanged performance after failure. Remaining paths may have less capacity or different congestion.

Business use

The business value can include more consistent infrastructure configuration and better use of compatible connected resources. Assess those benefits against operational complexity, interoperability, support arrangements and the workloads that must run.

Ask which resources are actually pooled, which configuration changes are supported, and what isolation exists between workloads. Request evidence for normal operation and degraded conditions rather than relying on a diagram showing many connections.

For AI workloads, distinguish communication-intensive training from independent inference jobs or data preparation. Their traffic patterns differ, so the same fabric can support them with different efficiency. Infrastructure choice should follow measured workload requirements, not the assumption that an AI label implies one universal network design.

People also ask

Is fabric computing necessarily cloud computing?

No. A fabric can support cloud infrastructure, but the connectivity and management approach can also exist in other environments.

Does it require one vendor?

Not as a conceptual requirement. Actual interoperability and supported configurations depend on the implementation.

Is every ordinary networked cluster a fabric-computing system?

Usage varies. The term commonly emphasizes coordinated resource connectivity and management, so describe the concrete architecture rather than relying only on the name.

Check your understanding

Six servers share a ten-gigabyte-per-second upstream path. Will doubling each server's local network-port speed necessarily double the aggregate data delivered through that path?

Answer: No. The shared upstream capacity remains a limit. End-to-end bottlenecks must be addressed before faster endpoint links can provide that benefit.

Related terms

• Data fabric
• Network fabric
• converged infrastructure
• composable infrastructure
• resource pooling
• bandwidth
• topology

Learn next

Read network topology and resource pooling. Draw the path between one workload and its data, then identify shared links, actual resource boundaries and what capacity remains after a selected component failure.

More in Data engineering and architecture

• Extensional Component - The extensional component of a database or knowledge system is its collection of represented instances or facts, considered separately from the definitions...
• External Schema - An external schema describes the database information presented to a particular user group or application.
• F-logic - F-logic, short for Frame Logic, is a formalism for representing and reasoning about objects, their properties, class membership and class relationships.
• Faceted classification - Faceted classification organizes information through several distinct characteristics, called facets, that can be combined to describe or retrieve an item.
• Fact table - A fact table is a table in a dimensional data model that records measurable business events, observations or relationships at a defined level of detail.
• Fail-stop - Fail-stop describes a failure behavior in which a component halts instead of continuing to produce externally visible erroneous results.

Sources

Gartner, Fabric-based Computing: https://www.gartner.com/en/information-technology/glossary/fabric-based-computing-fbc. Primary definition of the narrower modular architecture term.

Cisco, Unified Computing System Solution Overview: https://www.cisco.com/c/en/us/products/collateral/servers-unified-computing/solution-overview-c22-744677.html. Primary vendor description illustrating fabric connectivity and coordinated management; product-specific benefits are not generalised here. The capacity calculations and research-services case are original illustrations.

Previous: F-test of equality of variances Next: Face-to-face survey