@base <https://cloud.google.com/blog/products/networking/introducing-virgo-megascale-data-center-fabric> .
@prefix schema: <https://schema.org/> .
@prefix owl: <https://www.w3.org/2002/07/owl#> .

<#article> a schema:Article ;
  schema:headline "Introducing Virgo Network, Google’s scale-out AI data center fabric"@en ;
  schema:name "Introducing Virgo Network, Google’s scale-out AI data center fabric"@en ;
  schema:datePublished "2026-04-22" ;
  schema:inLanguage "en" ;
  schema:url <https://cloud.google.com/blog/products/networking/introducing-virgo-megascale-data-center-fabric> ;
  schema:publisher <#google-cloud-blog> ;
  schema:author <#benny-siman-tov>, <#arjun-singh> ;
  schema:about
    <#virgo-network>,
    <#campus-as-a-computer>,
    <#ai-hypercomputer>,
    <#scale-up-domain>,
    <#scale-out-fabric>,
    <#jupiter-front-end-network>,
    <#flat-two-layer-topology>,
    <#multi-planar-design>,
    <#bisection-bandwidth>,
    <#goodput>,
    <#fault-isolation>,
    <#deep-observability>,
    <#straggler-detection>,
    <#hang-detection>,
    <#mtbi>,
    <#mttr> ;
  schema:articleSection
    "Reimagining the data center network"@en,
    "Introducing Virgo Network: Megascale data center fabric"@en,
    "Improving reliability at scale"@en,
    "The foundation of the AI Hypercomputer"@en ;
  schema:abstract """The post introduces Virgo Network as Google’s new scale-out AI data center fabric, designed to provide massive scale, deterministic low latency, and reliability for modern AI workloads."""@en ;
  schema:articleBody """Google Cloud introduces Virgo Network as a specialized scale-out data center fabric built for the extreme requirements of AI training and serving workloads. The post argues that traditional general-purpose networks cannot meet modern AI needs for multi-data-center scale, rapidly growing bandwidth per accelerator, synchronized traffic bursts, and low latency. Google positions Virgo within a three-layer architecture consisting of a scale-up domain, a scale-out east-west accelerator fabric, and the Jupiter north-south front-end network. Virgo itself is described as a flat, two-layer non-blocking topology built on high-radix switches and multi-planar control domains, with claims of supporting 134,000 chips, up to 47 petabits per second of non-blocking bi-sectional bandwidth, 4x bandwidth per accelerator over the previous generation, and 40% lower unloaded fabric latency. The post also emphasizes reliability through fault isolation, sub-millisecond telemetry, and rapid detection of stragglers and hangs as the foundation for AI Hypercomputer resilience."""@en ;
  schema:hasPart
    <#part-constraints>,
    <#part-architecture>,
    <#part-virgo>,
    <#part-reliability>,
    <#part-foundation> ;
  schema:mentions
    <#defined-terms>,
    <#argument-howto>,
    <#faq-1>, <#faq-2>, <#faq-3>, <#faq-4>, <#faq-5>,
    <#faq-6>, <#faq-7>, <#faq-8>, <#faq-9>, <#faq-10>,
    <#google-cloud>,
    <#jupiter>,
    <#tpu-8t>,
    <#gemini-enterprise-agent-platform> .

<#google-cloud-blog> a schema:Blog ;
  schema:name "Google Cloud Blog"@en ;
  schema:url <https://cloud.google.com/blog/> ;
  schema:publisher <#google-cloud> .

<#google-cloud> a schema:Organization ;
  schema:name "Google Cloud"@en ;
  schema:url <https://cloud.google.com/> .

<#benny-siman-tov> a schema:Person ;
  schema:name "Benny Siman-Tov"@en ;
  schema:jobTitle "Senior Director Product Management"@en ;
  schema:affiliation <#google-cloud> .

<#arjun-singh> a schema:Person ;
  schema:name "Arjun Singh"@en ;
  schema:jobTitle "Engineering Fellow"@en ;
  schema:affiliation <#google-cloud> .

<#jupiter> a schema:SoftwareApplication, schema:Product ;
  schema:name "Jupiter"@en ;
  schema:brand <#google-cloud> ;
  schema:applicationCategory "Data center network"@en .

<#tpu-8t> a schema:Product ;
  schema:name "TPU 8t"@en ;
  schema:brand <#google-cloud> .

<#gemini-enterprise-agent-platform> a schema:SoftwareApplication, schema:Product ;
  schema:name "Gemini Enterprise Agent Platform"@en ;
  schema:brand <#google-cloud> .

<#virgo-network> a schema:SoftwareApplication, schema:Product ;
  schema:name "Virgo Network"@en ;
  schema:brand <#google-cloud> ;
  schema:applicationCategory "Scale-out AI data center fabric"@en ;
  schema:description """The post presents Virgo Network as Google’s new megascale AI fabric for scale-out east-west accelerator communication across pods and sites."""@en .

<#campus-as-a-computer> a schema:DefinedTerm ;
  schema:name "Campus-as-a-computer"@en ;
  schema:description """The design philosophy that treats a multi-data-center campus as a unified compute domain for very large-scale AI workloads."""@en .

<#ai-hypercomputer> a schema:DefinedTerm ;
  schema:name "AI Hypercomputer"@en ;
  schema:description """The broader Google infrastructure stack that Virgo Network underpins for large-scale training and serving workloads."""@en .

<#scale-up-domain> a schema:DefinedTerm ;
  schema:name "Scale-up domain"@en ;
  schema:description """The high-bandwidth, low-latency interconnect fabric within a single pod for tightly coupled accelerator communication."""@en .

<#scale-out-fabric> a schema:DefinedTerm ;
  schema:name "Scale-out accelerator fabric"@en ;
  schema:description """The east-west RDMA fabric optimized for massive horizontal scale across pods, deterministic latency, and resilience."""@en .

<#jupiter-front-end-network> a schema:DefinedTerm ;
  schema:name "Jupiter front-end network"@en ;
  schema:description """The north-south fabric that connects accelerators to storage and general-purpose compute and scales across sites for very large training runs."""@en .

<#flat-two-layer-topology> a schema:DefinedTerm ;
  schema:name "Flat two-layer non-blocking topology"@en ;
  schema:description """The Virgo network topology enabled by high-radix switches to reduce tiers and lower latency relative to traditional data center networks."""@en .

<#multi-planar-design> a schema:DefinedTerm ;
  schema:name "Multi-planar design"@en ;
  schema:description """The architecture of independent switching planes and control domains used to improve resilience and isolate localized failures."""@en .

<#bisection-bandwidth> a schema:DefinedTerm ;
  schema:name "Bi-sectional bandwidth"@en ;
  schema:description """The post’s central throughput metric for Virgo, used to express non-blocking scale across the fabric."""@en .

<#goodput> a schema:DefinedTerm ;
  schema:name "Goodput"@en ;
  schema:description """The effective workload throughput that Virgo is designed to maximize by reducing stragglers, congestion, and localized fault propagation."""@en .

<#fault-isolation> a schema:DefinedTerm ;
  schema:name "Fault isolation"@en ;
  schema:description """The hardware-level resilience strategy that keeps localized failures from degrading cluster-wide performance in synchronized training jobs."""@en .

<#deep-observability> a schema:DefinedTerm ;
  schema:name "Deep observability"@en ;
  schema:description """The use of sub-millisecond telemetry and high-fidelity visibility to monitor congestion, buffer behavior, and slowdown root causes at scale."""@en .

<#straggler-detection> a schema:DefinedTerm ;
  schema:name "Straggler detection"@en ;
  schema:description """The proactive localization of degraded nodes to prevent slow components from throttling synchronized AI workloads."""@en .

<#hang-detection> a schema:DefinedTerm ;
  schema:name "Hang detection"@en ;
  schema:description """The newly added capability to identify completely unresponsive nodes quickly and accelerate recovery."""@en .

<#mtbi> a schema:DefinedTerm ;
  schema:name "Mean-time between interruptions"@en ;
  schema:description """The reliability objective the post says Virgo improves by reducing disruptions in very large synchronized jobs."""@en .

<#mttr> a schema:DefinedTerm ;
  schema:name "Mean-time to recovery"@en ;
  schema:description """The recovery metric the post says Virgo and its software stack minimize by improving observability and failure localization."""@en .

<#defined-terms> a schema:DefinedTermSet ;
  schema:name "Defined terms for Introducing Virgo Network"@en ;
  schema:hasPart
    <#virgo-network>,
    <#campus-as-a-computer>,
    <#ai-hypercomputer>,
    <#scale-up-domain>,
    <#scale-out-fabric>,
    <#jupiter-front-end-network>,
    <#flat-two-layer-topology>,
    <#multi-planar-design>,
    <#bisection-bandwidth>,
    <#goodput>,
    <#fault-isolation>,
    <#deep-observability>,
    <#straggler-detection>,
    <#hang-detection>,
    <#mtbi>,
    <#mttr> ;
  schema:isPartOf <#article> .

<#part-constraints> a schema:WebPageElement ;
  schema:name "Why legacy networks break under AI"@en ;
  schema:position 1 ;
  schema:about <#campus-as-a-computer>, <#goodput> ;
  schema:text """The article says modern AI workloads impose four breaking constraints on legacy networks: multi-data-center scale, explosive bandwidth growth, synchronized bursts, and strict latency requirements."""@en .

<#part-architecture> a schema:WebPageElement ;
  schema:name "The three-layer architecture"@en ;
  schema:position 2 ;
  schema:about <#scale-up-domain>, <#scale-out-fabric>, <#jupiter-front-end-network> ;
  schema:text """Google splits the network into specialized scale-up, scale-out, and Jupiter front-end domains so each can evolve independently and optimize for its own role."""@en .

<#part-virgo> a schema:WebPageElement ;
  schema:name "What Virgo is"@en ;
  schema:position 3 ;
  schema:about <#virgo-network>, <#flat-two-layer-topology>, <#multi-planar-design>, <#bisection-bandwidth> ;
  schema:text """Virgo is described as a flat, two-layer, non-blocking, multi-planar scale-out fabric built on high-radix switches for lower latency and massive accelerator scale."""@en .

<#part-reliability> a schema:WebPageElement ;
  schema:name "Reliability at hundreds of thousands of chips"@en ;
  schema:position 4 ;
  schema:about <#fault-isolation>, <#deep-observability>, <#straggler-detection>, <#hang-detection>, <#mtbi>, <#mttr> ;
  schema:text """The post emphasizes independent switching planes, sub-millisecond telemetry, and automated localization of stragglers and hangs as the keys to protecting workload goodput at scale."""@en .

<#part-foundation> a schema:WebPageElement ;
  schema:name "Foundation of the AI Hypercomputer"@en ;
  schema:position 5 ;
  schema:about <#ai-hypercomputer> ;
  schema:text """The article concludes that Virgo provides the scale, predictable latency, and resilience needed to serve as the networking foundation for Google’s AI Hypercomputer."""@en .

<#argument-howto> a schema:HowTo ;
  schema:name "How the article builds the Virgo argument"@en ;
  schema:about <#virgo-network>, <#ai-hypercomputer> ;
  schema:isPartOf <#article> ;
  schema:step <#step-1>, <#step-2>, <#step-3>, <#step-4> ;
  schema:description """The article moves from AI-era constraints, to a specialized three-layer architecture, to Virgo’s fabric design, and finally to reliability and infrastructure implications."""@en .

<#step-1> a schema:HowToStep ;
  schema:name "Define the AI networking constraints"@en ;
  schema:position 1 ;
  schema:text "The post starts by naming the four pressures AI places on networks: scale, bandwidth, burstiness, and latency."@en ;
  schema:isPartOf <#argument-howto> .

<#step-2> a schema:HowToStep ;
  schema:name "Split the architecture into specialized layers"@en ;
  schema:position 2 ;
  schema:text "Google separates scale-up, scale-out, and north-south traffic into distinct domains to optimize each independently."@en ;
  schema:isPartOf <#argument-howto> .

<#step-3> a schema:HowToStep ;
  schema:name "Introduce Virgo as the scale-out fabric"@en ;
  schema:position 3 ;
  schema:text "Virgo is then positioned as the east-west accelerator fabric that delivers non-blocking scale, deterministic latency, and high resilience across pods."@en ;
  schema:isPartOf <#argument-howto> .

<#step-4> a schema:HowToStep ;
  schema:name "Tie performance to reliability and system design"@en ;
  schema:position 4 ;
  schema:text "The final move is to show that observability, fault isolation, and rapid mitigation are necessary for Virgo to function as the foundation of AI Hypercomputer."@en ;
  schema:isPartOf <#argument-howto> .

<#faq-1> a schema:Question ;
  schema:name "What is Virgo Network?"@en ;
  schema:text "What is Virgo Network?"@en ;
  schema:acceptedAnswer <#faq-1-answer> ;
  schema:isPartOf <#article> .
<#faq-1-answer> a schema:Answer ;
  schema:text "It is Google’s new scale-out AI data center fabric for east-west accelerator communication across pods in large AI training and serving systems."@en ;
  schema:isPartOf <#article> .

<#faq-2> a schema:Question ;
  schema:name "Why does Google say legacy networks are inadequate for modern AI?"@en ;
  schema:text "Why does Google say legacy networks are inadequate for modern AI?"@en ;
  schema:acceptedAnswer <#faq-2-answer> ;
  schema:isPartOf <#article> .
<#faq-2-answer> a schema:Answer ;
  schema:text "Because modern AI imposes extreme requirements for multi-data-center scale, rapidly growing bandwidth per accelerator, synchronized bursts, and low latency."@en ;
  schema:isPartOf <#article> .

<#faq-3> a schema:Question ;
  schema:name "What are the three layers in the reimagined architecture?"@en ;
  schema:text "What are the three layers in the reimagined architecture?"@en ;
  schema:acceptedAnswer <#faq-3-answer> ;
  schema:isPartOf <#article> .
<#faq-3-answer> a schema:Answer ;
  schema:text "The architecture consists of a scale-up domain, a scale-out accelerator fabric, and the Jupiter front-end north-south network."@en ;
  schema:isPartOf <#article> .

<#faq-4> a schema:Question ;
  schema:name "What makes Virgo’s topology different?"@en ;
  schema:text "What makes Virgo’s topology different?"@en ;
  schema:acceptedAnswer <#faq-4-answer> ;
  schema:isPartOf <#article> .
<#faq-4-answer> a schema:Answer ;
  schema:text "The post says Virgo uses high-radix switches to create a flat, two-layer, non-blocking topology with lower latency than traditional multi-tier data center networks."@en ;
  schema:isPartOf <#article> .

<#faq-5> a schema:Question ;
  schema:name "What scale claims does Google make for Virgo?"@en ;
  schema:text "What scale claims does Google make for Virgo?"@en ;
  schema:acceptedAnswer <#faq-5-answer> ;
  schema:isPartOf <#article> .
<#faq-5-answer> a schema:Answer ;
  schema:text "The article says Virgo can link 134,000 TPU 8t chips with up to 47 petabits per second of non-blocking bi-sectional bandwidth in a single fabric."@en ;
  schema:isPartOf <#article> .

<#faq-6> a schema:Question ;
  schema:name "What performance improvements are claimed?"@en ;
  schema:text "What performance improvements are claimed?"@en ;
  schema:acceptedAnswer <#faq-6-answer> ;
  schema:isPartOf <#article> .
<#faq-6-answer> a schema:Answer ;
  schema:text "Google claims up to 4x bandwidth per accelerator over the previous generation and 40% lower unloaded fabric latency for TPUs."@en ;
  schema:isPartOf <#article> .

<#faq-7> a schema:Question ;
  schema:name "Why is fault isolation central to the design?"@en ;
  schema:text "Why is fault isolation central to the design?"@en ;
  schema:acceptedAnswer <#faq-7-answer> ;
  schema:isPartOf <#article> .
<#faq-7-answer> a schema:Answer ;
  schema:text "Because in synchronized training jobs a single localized hardware failure can degrade cluster-wide goodput unless failures are isolated quickly."@en ;
  schema:isPartOf <#article> .

<#faq-8> a schema:Question ;
  schema:name "What role does observability play?"@en ;
  schema:text "What role does observability play?"@en ;
  schema:acceptedAnswer <#faq-8-answer> ;
  schema:isPartOf <#article> .
<#faq-8-answer> a schema:Answer ;
  schema:text "Sub-millisecond telemetry is used to detect congestion, optimize buffers, and pinpoint the root causes of slowdowns across hardware and software."@en ;
  schema:isPartOf <#article> .

<#faq-9> a schema:Question ;
  schema:name "What are stragglers and hangs in this context?"@en ;
  schema:text "What are stragglers and hangs in this context?"@en ;
  schema:acceptedAnswer <#faq-9-answer> ;
  schema:isPartOf <#article> .
<#faq-9-answer> a schema:Answer ;
  schema:text "Stragglers are degraded nodes and hangs are unresponsive nodes; both can throttle synchronized AI workloads if not localized rapidly."@en ;
  schema:isPartOf <#article> .

<#faq-10> a schema:Question ;
  schema:name "How does the post position Virgo relative to AI Hypercomputer?"@en ;
  schema:text "How does the post position Virgo relative to AI Hypercomputer?"@en ;
  schema:acceptedAnswer <#faq-10-answer> ;
  schema:isPartOf <#article> .
<#faq-10-answer> a schema:Answer ;
  schema:text "It is positioned as the networking foundation of Google’s AI Hypercomputer, providing the scale, latency, and reliability needed for modern AI workloads."@en ;
  schema:isPartOf <#article> .