Understand the system
How it works
AI infrastructure succeeds when data and synchronization arrive on time; network, storage, and software stalls are stranded silicon in another form.
Connecting the scale-out network
Select a capability to explore its role and connections.
Selected: Switching silicon
Switching silicon
Switch ASICs determine port density, throughput, buffering, routing features, power, and supported network topologies.
Explore Switching silicon in AtlasDocumented connections
- Switching silicon → Optical transceivers
Switch ports drive optical modules where electrical reach is insufficient.
Selected documented dependencies. Arrows retain their Atlas direction; they do not represent quantities or a complete engineering process.
Large AI systems are data-movement machines. Scale-up links coordinate devices within a node or rack, scale-out fabrics connect thousands more, optics carry signals across distance, and storage streams training data while absorbing checkpoints. Above them, drivers, compilers, kernels, collectives, schedulers, frameworks, and observability software decide whether the physical system behaves like one dependable computer.
Scale-up and scale-out
Scale-up fabrics prioritize tightly coupled accelerator communication; scale-out networks connect racks and pods. The design must match collective patterns, failure domains, and oversubscription targets.
- Distinguish electrical and optical reaches
- Map topology to model parallelism
Supporting evidence · 3
- [1] UALink 1.0 defines 200 Gb/s-per-lane links and a scale-up fabric designed for as many as 1,024 accelerators.
- [2] Broadcom specifies 102.4 Tb/s of switching capacity for a single Tomahawk 6 Ethernet switch chip.
- [3] NVIDIA reported Data Center networking revenue of $12.990 billion in fiscal 2025 and $31.376 billion in fiscal 2026, an increase of 142 percent; this is one supplier's audited revenue rather than total network capital expenditure.
Optics is the distance tax
SerDes, copper cables, transceivers, fibers, connectors, and switches convert and regenerate signals across the cluster. Power, heat, insertion loss, and manufacturing yield rise with bandwidth.
- Reach determines where electrical links become optical
- Transceivers are both power loads and supply-chain nodes
Storage protects expensive time
Training pipelines need parallel ingest, metadata performance, checkpoint writes, and rapid restore. A storage pause idles the costliest resources in the building.
- Trace dataset, checkpoint, and recovery paths
- Measure accelerator idle time caused by I/O
Supporting evidence · 3
- [1] Industry presenters working through SNIA describe AI storage as needing to sustain dataset ingest while absorbing checkpoints and supporting recovery without leaving accelerators idle.
- [2] CXL 3.1 supports fabric-attached memory use cases including memory sharing and pooling across hosts and devices.
- [3] NVM Express Base Specification 2.4 defines the base architecture and host-controller interface for NVM subsystems; it does not prescribe AI workload throughput.
Software is the utilization layer
Compilers, optimized kernels, runtimes, collectives, parallelism strategies, schedulers, and monitoring translate hardware into useful work. Portability and tuning depth shape both economics and resilience.
- Surface lock-in without assuming equivalence
- Tie software versions to measured utilization
Supporting evidence · 2
Featured evidence
Evidence in context
Facility-stated storage capacity references
Published capacity figures for Perlmutter all-flash scratch, Aurora DAOS, and the Orion filesystem capacity tier.View chart values
| Category | Facility-stated capacity |
|---|---|
| Perlmutter | 44 PB |
| Aurora DAOS | 230 PB |
| Orion | 679 PB |
Key indicators
Key measures & constraints
- 01
Broadcom specifies 102.4 Tb/s of switching capacity for a single Tomahawk 6 Ethernet switch chip.
- Class
- vendor spec
- Geography
- Global
- Period
- Tomahawk 6 generation
- Confidence
- high
- 02
PyTorch documents data, fully sharded data, tensor, and pipeline parallelism as complementary approaches for distributed model training.
- Class
- fact
- Geography
- Global
- Period
- 2025 documentation
- Confidence
- high
- 03
U.S. laboratory documentation states 44 PB for Perlmutter scratch, 230 PB for Aurora DAOS, and a 679 PB usable capacity tier for Orion; these differ in architecture and lifecycle policy and are not interchangeable workload benchmarks.
- Class
- fact
- Geography
- United States
- Period
- Live facility documentation reviewed August 2026
- Confidence
- high
- 04
NVIDIA reported Data Center networking revenue of $12.990 billion in fiscal 2025 and $31.376 billion in fiscal 2026, an increase of 142 percent; this is one supplier's audited revenue rather than total network capital expenditure.
- Class
- fact
- Geography
- Global
- Period
- NVIDIA fiscal years 2025 and 2026
- Confidence
- high
System map
Connections across the system
Explore inputs, outputs, and shared capabilities in the Atlas. Open the register for every documented relationship.
Enters this system 1
- AI rack system→ Scale-out networkTrace relationship
Leaves this system 3
- AI rack system← Runtime & frameworksTrace relationship
- Accelerator architecture← Runtime & frameworksTrace relationship
Shared capabilities can belong to more than one System. These links carry materials, energy, information, capital, or permissions; they describe dependencies, not quantities or a complete process model.
Explore 10 capabilities and operating boundaries
All documented relationships · 15
- criticaloutput
Drivers, kernels, frameworks, and scheduling convert hardware into useful work.
Trace relationshipSupporting evidence · 2
- criticalinternal
Endpoint adapters and switching silicon form the electrical endpoints and forwarding core of the fabric.
Trace relationship - importantoutput
Software capabilities and workload behavior feed back into hardware design.
Trace relationship - importantoutput
Scale-up links couple accelerators within the rack-scale computer.
Trace relationship - importantinput
Rack endpoints attach to the cluster's scale-out switching fabric.
Trace relationshipSupporting evidence · 2
- importantinternal
Optical transceivers and fiber carry high-rate links beyond copper reach.
Trace relationship - importantinternal
Storage traffic reaches compute through the data-center fabric.
Trace relationship - importantinternal
Parallelism and collective libraries shape network traffic and topology needs.
Trace relationship - importantinternal
Datasets and checkpoints feed training and recovery workflows.
Trace relationship - importantinternal
NICs, SmartNICs & DPUs supplies a distinct operating capability within its broader Atlas system.
Trace relationship - importantinternal
Switching silicon supplies a distinct operating capability within its broader Atlas system.
Trace relationshipSupporting evidence · 2
- [1] Broadcom specifies 102.4 Tb/s of switching capacity for a single Tomahawk 6 Ethernet switch chip.
- [2] NVIDIA reported Data Center networking revenue of $12.990 billion in fiscal 2025 and $31.376 billion in fiscal 2026, an increase of 142 percent; this is one supplier's audited revenue rather than total network capital expenditure.
- importantinternal
Optical transceivers supplies a distinct operating capability within its broader Atlas system.
Trace relationship - importantinternal
Fiber & cabling supplies a distinct operating capability within its broader Atlas system.
Trace relationship - importantinternal
Storage media & controllers supplies a distinct operating capability within its broader Atlas system.
Trace relationshipSupporting evidence · 2
- [1] NVM Express Base Specification 2.4 defines the base architecture and host-controller interface for NVM subsystems; it does not prescribe AI workload throughput.
- [2] U.S. laboratory documentation states 44 PB for Perlmutter scratch, 230 PB for Aurora DAOS, and a 679 PB usable capacity tier for Orion; these differ in architecture and lifecycle policy and are not interchangeable workload benchmarks.
- importantinternal
Switch ports drive optical modules where electrical reach is insufficient.
Trace relationship
Named companies + institutions
Organizations & their roles
- principal · public company
NVIDIA
NVIDIA CorporationComputing-platform company that designs AI accelerators, interconnects, systems, and the software stack that operates them.
Roles, locations & evidence
- Role
- supplier
- Products / service
- Data-center networking platforms · CUDA software platform
- Documented activity
- World
- Headquarters
- United States
Why this organization belongs in the System- Scale-out network
NVIDIA's filing-backed operating portfolio includes data-center networking platforms, and the attached quantitative operating or investment measure supports principal status at this Atlas boundary; no subregional operating place is asserted.
- Runtime & frameworks
NVIDIA's filing-backed operating portfolio includes cuda software platform; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.
- principal · standards body
NVM Express
NVM Express, Inc.Standards organization maintaining the NVMe storage interface specifications.
Roles, locations & evidence
- Role
- standards setter
- Products / service
- NVMe storage interface standards
- Documented activity
- World
- Headquarters
- Not cataloged
Why this organization belongs in the System- AI storage
Primary-source evidence connects this institution to nvme storage interface standards in the specified Atlas capability.
- principal · consortium
UALink Consortium
Industry consortium developing a scale-up interconnect specification for accelerator systems.
Roles, locations & evidence
- Role
- standards setter
- Products / service
- accelerator scale-up interconnect specification
- Documented activity
- World
- Headquarters
- Not cataloged
Why this organization belongs in the System- Scale-up fabric
Primary-source evidence connects this institution to accelerator scale-up interconnect specification in the specified Atlas capability.
- material · public company
Arista Networks
Arista Networks, Inc.Cloud-networking company supplying high-speed switching systems, routing platforms, and network software.
Roles, locations & evidence
- Role
- supplier
- Products / service
- Network operating software · Switch platforms built around merchant and custom silicon · Data-center switching systems
- Documented activity
- World
- Headquarters
- United States
Why this organization belongs in the System- Runtime & frameworks
Arista Networks's filing-backed operating portfolio includes network operating software; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.
- Switching silicon
Arista Networks's filing-backed operating portfolio includes switch platforms built around merchant and custom silicon; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.
- Scale-out network
Arista Networks's filing-backed operating portfolio includes data-center switching systems; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.
All other organizations · 16
- material · public company
Broadcom
Broadcom Inc.Semiconductor and infrastructure-software company supplying switching, connectivity, custom silicon, storage, and optical components.
Roles, locations & evidence
- Role
- supplier · designer
- Products / service
- Storage connectivity and controller silicon · Network adapters and connectivity silicon · Ethernet switching silicon
- Documented activity
- World
- Headquarters
- United States
Why this organization belongs in the System- Storage media & controllers
Broadcom's filing-backed operating portfolio includes storage connectivity and controller silicon; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.
- NICs, SmartNICs & DPUs
Broadcom's filing-backed operating portfolio includes network adapters and connectivity silicon; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.
- Switching silicon
Broadcom's filing-backed operating portfolio includes ethernet switching silicon; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.
- material · public company
Cisco
Cisco Systems, Inc.Networking and security company supplying switching, routing, optical, observability, and data-center infrastructure.
Roles, locations & evidence
- Role
- supplier
- Products / service
- Optical networking systems and modules · Network operating and observability software · Data-center switching and routing
- Documented activity
- World
- Headquarters
- United States
Why this organization belongs in the System- Optical transceivers
Cisco's filing-backed operating portfolio includes optical networking systems and modules; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.
- Runtime & frameworks
Cisco's filing-backed operating portfolio includes network operating and observability software; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.
- Scale-out network
Cisco's filing-backed operating portfolio includes data-center switching and routing; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.
- material · public company
Equinix
Equinix, Inc.Digital-infrastructure company operating carrier-neutral data centers and interconnection services across global markets.
Roles, locations & evidence
- Role
- operator
- Products / service
- Digital interconnection services
- Documented activity
- World
- Headquarters
- United States
Why this organization belongs in the System- Scale-out network
Equinix's filing-backed operating portfolio includes digital interconnection services; the listed places are explicit role geographies, and headquarters is not used as a proxy. Representative status does not imply market rank.
- material · public company
HPE
Hewlett Packard Enterprise CompanyEnterprise infrastructure company supplying servers, supercomputers, networking, storage, and services.
Roles, locations & evidence
- Role
- supplier
- Products / service
- Enterprise data storage
- Documented activity
- World
- Headquarters
- United States
Why this organization belongs in the System- AI storage
HPE's filing-backed operating portfolio includes enterprise data storage; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.
- material · public company
Marvell
Marvell Technology, Inc.Data-infrastructure semiconductor company supplying networking, storage, interconnect, and custom compute silicon.
Roles, locations & evidence
- Role
- supplier · designer
- Products / service
- Storage controller silicon · Switching and connectivity silicon · Electro-optical DSP silicon
- Documented activity
- World
- Headquarters
- United States
Why this organization belongs in the System- Storage media & controllers
Marvell's filing-backed operating portfolio includes storage controller silicon; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.
- Switching silicon
Marvell's filing-backed operating portfolio includes switching and connectivity silicon; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.
- Optical transceivers
Marvell's filing-backed operating portfolio includes electro-optical dsp silicon; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.
- material · public company
Micron
Micron Technology, Inc.Memory and storage manufacturer supplying DRAM, HBM, NAND, and related products for AI systems.
Roles, locations & evidence
- Role
- supplier
- Products / service
- NAND and SSD products
- Documented activity
- World
- Headquarters
- Boise
Why this organization belongs in the System- Storage media & controllers
Micron's filing-backed operating portfolio includes nand and ssd products; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.
- material · nonprofit
PyTorch Foundation
Open-source foundation supporting the PyTorch machine-learning framework and distributed-computing ecosystem.
Roles, locations & evidence
- Role
- standards setter
- Products / service
- distributed machine-learning software
- Documented activity
- World
- Headquarters
- Not cataloged
Why this organization belongs in the System- Runtime & frameworks
Primary-source evidence connects this institution to distributed machine-learning software in the specified Atlas capability.
- material · standards body
SNIA
Storage Networking Industry AssociationIndustry association developing storage standards, education, and technical guidance.
Roles, locations & evidence
- Role
- standards setter
- Products / service
- AI storage technical guidance
- Documented activity
- World
- Headquarters
- Not cataloged
Why this organization belongs in the System- AI storage
Primary-source evidence connects this institution to ai storage technical guidance in the specified Atlas capability.
- representative · public company
AMD
Advanced Micro Devices, Inc.Semiconductor company that designs CPUs, GPUs, adaptive computing products, and associated AI software.
Roles, locations & evidence
- Role
- supplier
- Products / service
- ROCm software platform
- Documented activity
- World
- Headquarters
- United States
Why this organization belongs in the System- Runtime & frameworks
AMD's filing-backed operating portfolio includes rocm software platform; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.
- representative · public company
Coherent
Coherent Corp.Photonics company supplying lasers, optical communications components, transceivers, and engineered materials.
Roles, locations & evidence
- Role
- manufacturer · supplier
- Products / service
- Optical transceivers and communications components · Optical communications components
- Documented activity
- World
- Headquarters
- United States
Why this organization belongs in the System- Optical transceivers
Coherent's filing-backed operating portfolio includes optical transceivers and communications components; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.
- Optics & fiber
Coherent's filing-backed operating portfolio includes optical communications components; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.
- representative · public company
Corning
Corning IncorporatedMaterials-science company supplying optical fiber, cable, connectivity products, glass, and ceramic technologies.
Roles, locations & evidence
- Role
- manufacturer · supplier
- Products / service
- Optical fiber and cabling · Data-center optical connectivity
- Documented activity
- World
- Headquarters
- United States
Why this organization belongs in the System- Fiber & cabling
Corning's filing-backed operating portfolio includes optical fiber and cabling; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.
- Optics & fiber
Corning's filing-backed operating portfolio includes data-center optical connectivity; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.
- representative · public company
Dell Technologies
Dell Technologies Inc.Enterprise technology company supplying servers, storage, networking, and integrated AI infrastructure.
Roles, locations & evidence
- Role
- supplier
- Products / service
- Enterprise storage systems
- Documented activity
- World
- Headquarters
- United States
Why this organization belongs in the System- AI storage
Dell Technologies's filing-backed operating portfolio includes enterprise storage systems; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.
- representative · public company
Lumentum
Lumentum Holdings Inc.Optical and photonic-products company supplying components used in cloud, telecom, and industrial laser systems.
Roles, locations & evidence
- Role
- manufacturer · supplier
- Products / service
- Optical communications components · Photonic components for high-speed links
- Documented activity
- World
- Headquarters
- United States
Why this organization belongs in the System- Optical transceivers
Lumentum's filing-backed operating portfolio includes optical communications components; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.
- Optics & fiber
Lumentum's filing-backed operating portfolio includes photonic components for high-speed links; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.
- representative · public company
NetApp
NetApp, Inc.Data-infrastructure company supplying file, block, object, and cloud storage systems and software.
Roles, locations & evidence
- Role
- supplier
- Products / service
- Enterprise and cloud data storage
- Documented activity
- World
- Headquarters
- United States
Why this organization belongs in the System- AI storage
NetApp's filing-backed operating portfolio includes enterprise and cloud data storage; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.
- representative · public company
Seagate
Seagate Technology Holdings plcMass-capacity-storage company supplying hard-disk drives and storage systems for cloud and enterprise deployments.
Roles, locations & evidence
- Role
- supplier · manufacturer
- Products / service
- Cloud and enterprise storage media · Mass-capacity hard-disk drives
- Documented activity
- World
- Headquarters
- Ireland
Why this organization belongs in the System- AI storage
Seagate's filing-backed operating portfolio includes cloud and enterprise storage media; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.
- Storage media & controllers
Seagate's filing-backed operating portfolio includes mass-capacity hard-disk drives; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.
- representative · public company
Western Digital
Western Digital CorporationData-storage company supplying hard-disk drives and related technologies for cloud and enterprise infrastructure.
Roles, locations & evidence
- Role
- supplier · manufacturer
- Products / service
- Capacity storage media · Hard-disk drives and storage technologies · Vertically integrated HDD assembly and test
- Documented activity
- World · Malaysia · Thailand · Philippines
- Headquarters
- United States
Why this organization belongs in the System- AI storage
Western Digital's filing-backed operating portfolio includes capacity storage media; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.
- Storage media & controllers
Western Digital's filing-backed operating portfolio includes hard-disk drives and storage technologies; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.
- Storage media & controllers
Western Digital's fiscal 2026 audited filing directly locates its HDD assembly-and-test operations in all three countries. Principal status is withheld because this edition does not close a concentration, capacity, shipment, revenue, or installed-base measure for the role.
Read this System in the field guide
Continue in the Volumes
Start with three selected readings, or explore the complete reading list.
Complete reading list
Fabric
15 spreads- V · 01The real workloadOpen spread
- V · 02Distance has a priceOpen spread
- V · 03Scale upOpen spread
- V · 04Scale outOpen spread
- V · 05EndpointsOpen spread
- V · 06Copper and lightOpen spread
- V · 07Routes and patchingOpen spread
- V · 08Congestion and topologyOpen spread
- V · 09The data pipelineOpen spread
- V · 10Checkpoint and restoreOpen spread
- V · 11Memory as fabricOpen spread
- V · 12Observe and recoverOpen spread
- V · 13Programming stacksOpen spread
- V · 14ParallelismOpen spread
- V · 15Useful workOpen spread
Evidence & review
currentEvidence health29 active claims · next review Oct 13, 2026+
- Active claims
- 29
- Sources
- 48
- High volatility
- 4
- Next review
- Oct 13, 2026
- Due soon
- 0
- Overdue
- 0
- Superseded
- 0