Predictika · Cisco Nexus AI cluster fabric design

An AI agent's GPU cluster fabric design, checked before it reaches the simulator

An SE, a partner or an AI agent proposes the east-west fabric for a GPU cluster: which Cisco Nexus 9000 leaf and spine switches, how many spines, which optic, how far the links run, which NX-OS release, whether Spectrum-X is on and which license covers it. Before the design goes to a simulator, a lab or an order, the proposal goes to the Predictika validator. It recomputes the GPU count, the leaf count and the oversubscription from the nodes and spines, checks every optic against the ports and the distance, checks the Spectrum-X release and license, compares the figures the proposal states with the computed ones, and returns every reason in one answer. Hard rules block; soft recommendations route to a person for approval. This checks that a design is legal, not how it performs: it does not simulate traffic. Nothing here is priced. This is an illustrative model built from Cisco's public documentation. Cisco Systems is not involved with it and has not endorsed it.

Scope: two-tier leaf-spine east-west fabrics for GPU clusters on Cisco Nexus 9000 switches: N9364E-SG2-O, N9164E-NS4-O and N9332D-GX2B. See everything this validator covers

This validator checks one kind of proposal: a two-tier leaf-spine east-west fabric for an eight-GPU-node cluster built from Cisco Nexus 9000 switches, as in Cisco's Nexus 9000 AI-era reference architecture. It handles any proposal that picks from these switches, optics and NX-OS releases, not only the scenarios on this page. It checks whether the design is legal before it is simulated, tested or ordered. It does not simulate traffic or predict performance, and it does not cover storage or north-south networks, configuration or pricing.

  • Covered: sizing — GPUs from the node count (8 GPUs per node, one 400G fabric link per GPU), leaf switches from the node-facing ports of the chosen leaf, uplinks and spines needed for 1:1, and the oversubscription of what is proposed. Two-to-one blocks; anything above 1:1 and below 2:1 needs approval. Counts are in 400G equivalents.
  • Covered: optics — form factor (OSFP, QSFP-DD or QSFP) against the ports of both the leaf and the spine, and the reach of the optic against the leaf-to-spine distance, for OSFP-800G-DR8, QDD-400G-DR4 and QSFP-200G-SR4.
  • Covered: Spectrum-X — the NX-OS 10.6(1)F minimum on both leaf and spine, and the DCN Advantage license; where the data sheet does not require the license (the N9164E-NS4-O), the question goes to approval.
  • Covered: the reference architecture — the 128-node (1,024 GPU) ceiling, the 4-node scalable unit, and whether each switch is a leaf or spine model in the reference; others route to approval. Load balancing and lossless configuration are checked too.
  • Covered: the proposal itself — the GPU count and oversubscription the proposal states are compared with the computed figures; a mismatch blocks.
  • Not covered — traffic simulation, cabling and breakout, power and cooling, three-tier fabrics with superspines, storage and north-south networks, switch configuration, and any price or availability. Other areas can be added.
  • Only stated rules are modeled — Cisco does not publish a minimum NX-OS release per optic, so none is checked. Optic reaches are the standard figures for the optic type.
The AI design agentProposes the GPU back-end fabric from the customer's requirements and says it is ready for the build team.
Predictika validatorChecks every rule at once and returns the verdict, each finding, and the computed figures.

8 scenarios in 4 groups · each group runs from the agent’s first attempt to the outcome · click one to switch

What the agent proposes

The AI design agent says

Illustrative agent statement. The validator receives the structured request below.

Proposing agent: AI design agent (scripted)

Before sending, the agent must call Recorded run

Predictika validator


      

        
        
        
      

      
Waiting for the call.

      
    

  

  
  

    
Where rules-based CPQ gets it wrong
Click to expand
    
Our assessment of what a typical rules-based CPQ (product rules, price rules and validation rules, as in Salesforce CPQ) does with the same cases. It is a judgment about effort and reliability, not a claim that nothing can be built.

    
What the validator caughtTypical rules-based CPQPredictika
Fabric sizing that derives leaves, uplinks and spines from the node countOnly with workarounds
The leaf count depends on the node count in steps and on the port count of the chosen leaf, and the spine count on the port count of the chosen spine, so each is a formula that has to be rewritten for every switch model, and a 3-spine proposal for 1,024 GPUs passes unless someone adds the arithmetic.		The leaf, uplink and spine counts are computed on every call and compared with what was proposed; 3 spines for 1,024 GPUs comes back as 2.67 to 1 oversubscribed, with the 8 spines that 1:1 needs, and the stated 1:1 is caught.
Optics against distance and form factorOnly with care
Reach and form factor are compatibility rules per optic and per switch, on both ends of the link. They can be written, but every new optic multiplies them, and a distance has to be compared with a reach.		Each optic carries its form factor and reach; a 400G QSFP-DD optic in OSFP ports and a 700 metre run on a 500 metre optic are both caught on the same call.
Feature rules that depend on release and license togetherOnly with workarounds
Spectrum-X needs NX-OS 10.6(1)F on both tiers and the DCN Advantage license, except where the data sheet says otherwise. That is a release rule and a license rule that interact, and a platform exception.		The release is checked on the leaf and on the spine separately, the license is checked, and the N9164E-NS4-O exception routes to approval rather than a false block or a false pass.
Hard rules versus recommendations, and figures the proposal itself statesOnly with workarounds
Approval chains handle one threshold well, but here some findings block, some only need a person, and the proposal's own stated GPU count and oversubscription must be checked against the computed ones.		Each finding is labeled block or approval, and the stated figures travel with the proposal and are compared on every call.

  

  

    
What this validator catches: findings from its test cases
Click to expand
    
Every rule is checked on every call, in any order the choices arrive; one problem never hides another. Blocking findings stop the design; approval findings need a person to sign off; everything else is computed. The wording shown is an example from the test runs; the numbers change with the proposal.

    
1Fabric is oversubscribed (blocks): The fabric is 2.67 to 1 oversubscribed: 1024 leaf-to-spine uplinks are needed for 1:1 but 3 x N9364E-SG2-O give 384 ports. 1:1 needs 8 spines.
2Optic cannot reach the distance (blocks): The leaf-to-spine run is 700 m but QDD-400G-DR4 reaches 500 m over single-mode fiber.
3Optic does not fit the switch ports (blocks): Optic QDD-400G-DR4 has the QSFP-DD form factor, but leaf N9364E-SG2-O has OSFP ports and spine N9364E-SG2-O has OSFP ports; the optic must fit both.
4Spectrum-X needs a newer NX-OS on the leaf (blocks): Spectrum-X needs NX-OS 10.6(1)F or later (Cisco data sheet); leaf N9364E-SG2-O is set to run 10.5(3)F.
5Spectrum-X needs a newer NX-OS on the spine (blocks): Spectrum-X needs NX-OS 10.6(1)F or later (Cisco data sheet); spine N9364E-SG2-O is set to run 10.5(3)F.
6Spectrum-X needs the DCN Advantage license (blocks): Spectrum-X needs the DCN Advantage license (Cisco data sheet); the proposal includes DCN Essentials.
7Stated GPU count does not match (blocks): Stated GPU count does not match: the agent stated 1024; the validator computes 512 (64 nodes x 8 GPUs).
8Stated oversubscription does not match (blocks): Stated oversubscription does not match: the agent stated 1.0 to 1; the validator computes 2.67 to 1.
9DCN license to confirm with Cisco (needs approval): The proposal includes DCN Essentials with Spectrum-X on a design with leaf N9164E-NS4-O and spine N9164E-NS4-O. The Cisco data sheet requires DCN Advantage for Spectrum-X but does not require the Spectrum-X license for the N9164E-NS4-O; confirm the license with Cisco.
10Larger than the reference architecture (needs approval): 144 nodes is above the 128 nodes (1,024 GPUs) of the Cisco reference architecture; confirm the design with Cisco.
11No lossless configuration (needs approval): No lossless configuration proposed; RoCE east-west fabrics are normally run lossless with PFC and ECN.
12Oversubscribed against the 1:1 target (needs approval): The fabric is 1.6 to 1 oversubscribed; AI east-west fabrics are normally built at 1:1, so a design below 2:1 but above 1:1 needs a person to accept it.
13Partial scalable unit (needs approval): The Cisco reference architecture grows in 4-node scalable units; 30 nodes leaves a partial unit of 2.
14Plain ECMP load balancing (needs approval): Plain ECMP proposed instead of Spectrum-X with fine-grain load balancing; Cisco positions Spectrum-X for AI east-west fabrics.
15Switch is not a listed leaf (needs approval): N9364E-SG2-Q is not one of the leaf models in the Cisco Nexus 9000 AI reference architecture this validator covers.
16Switch is not a listed spine (needs approval): N9332D-GX2B is not one of the spine models in the Cisco Nexus 9000 AI reference architecture this validator covers.

  

  

    
Scenarios and results
Click to expand
    
The recorded result of the validator for each scenario on this page: the request the AI agent sent, the verdict, and every finding, in the order they appear above.

    
1. Sizing, optics and release: 3 spines, 400G optics over 700 m, Spectrum-X on 10.5(3)F

    
The agent's request. Please check the east-west GPU fabric for Sable Ridge AI Cloud: 128 GPU nodes, 1,024 GPUs in all, on N9364E-SG2-O leaves running NX-OS 10.5(3)F and 3 N9364E-SG2-O spines running NX-OS 10.6(1)F, linked with QDD-400G-DR4 optics over 700 metres, with Spectrum-X and the DCN Advantage license, PFC and ECN lossless, stated as 1.0 to 1. It is ready for the build team.

    
Verdict: Blocked.

    
    
      
  • Fabric is oversubscribed. The fabric is 2.67 to 1 oversubscribed: 1024 leaf-to-spine uplinks are needed for 1:1 but 3 x N9364E-SG2-O give 384 ports. 1:1 needs 8 spines.
    • 
      
  • Optic does not fit the switch ports. Optic QDD-400G-DR4 has the QSFP-DD form factor, but leaf N9364E-SG2-O has OSFP ports and spine N9364E-SG2-O has OSFP ports; the optic must fit both.
    • 
      
  • Optic cannot reach the distance. The leaf-to-spine run is 700 m but QDD-400G-DR4 reaches 500 m over single-mode fiber.
    • 
      
  • Spectrum-X needs a newer NX-OS on the leaf. Spectrum-X needs NX-OS 10.6(1)F or later (Cisco data sheet); leaf N9364E-SG2-O is set to run 10.5(3)F.
    • 
      
  • Stated oversubscription does not match. Stated oversubscription does not match: the agent stated 1.0 to 1; the validator computes 2.67 to 1.
    • 
    

    
2. Sizing, optics and release: 800G optics and 10.6(1)F, still 3 spines

    
The agent's request. Please check the east-west GPU fabric for Sable Ridge AI Cloud: 128 GPU nodes, 1,024 GPUs in all, on N9364E-SG2-O leaves running NX-OS 10.6(1)F and 3 N9364E-SG2-O spines running NX-OS 10.6(1)F, linked with OSFP-800G-DR8 optics over 700 metres, with Spectrum-X and the DCN Advantage license, PFC and ECN lossless, stated as 1.0 to 1. It is ready for the build team.

    
Verdict: Blocked.

    
    
      
  • Fabric is oversubscribed. The fabric is 2.67 to 1 oversubscribed: 1024 leaf-to-spine uplinks are needed for 1:1 but 3 x N9364E-SG2-O give 384 ports. 1:1 needs 8 spines.
    • 
      
  • Optic cannot reach the distance. The leaf-to-spine run is 700 m but OSFP-800G-DR8 reaches 500 m over single-mode fiber.
    • 
      
  • Stated oversubscription does not match. Stated oversubscription does not match: the agent stated 1.0 to 1; the validator computes 2.67 to 1.
    • 
    

    
3. Sizing, optics and release: 8 spines, 400 m run

    
The agent's request. Please check the east-west GPU fabric for Sable Ridge AI Cloud: 128 GPU nodes, 1,024 GPUs in all, on N9364E-SG2-O leaves running NX-OS 10.6(1)F and 8 N9364E-SG2-O spines running NX-OS 10.6(1)F, linked with OSFP-800G-DR8 optics over 400 metres, with Spectrum-X and the DCN Advantage license, PFC and ECN lossless, stated as 1.0 to 1. It is ready for the build team.

    
Verdict: Allowed.

    
4. Scalable units and license: 30 nodes, DCN Essentials with Spectrum-X

    
The agent's request. Please check the east-west GPU fabric for Altair AI Lab: 30 GPU nodes, 240 GPUs in all, on N9164E-NS4-O leaves running NX-OS 10.6(1)F and 2 N9164E-NS4-O spines running NX-OS 10.6(1)F, linked with OSFP-800G-DR8 optics over 200 metres, with Spectrum-X and the DCN Essentials license, PFC and ECN lossless, stated as 1.0 to 1. It is ready for the build team.

    
Verdict: Needs approval.

    
    
      
  • Partial scalable unit. The Cisco reference architecture grows in 4-node scalable units; 30 nodes leaves a partial unit of 2.
    • 
      
  • DCN license to confirm with Cisco. The proposal includes DCN Essentials with Spectrum-X on a design with leaf N9164E-NS4-O and spine N9164E-NS4-O. The Cisco data sheet requires DCN Advantage for Spectrum-X but does not require the Spectrum-X license for the N9164E-NS4-O; confirm the license with Cisco.
    • 
    

    
5. Scalable units and license: 32 nodes, DCN Advantage

    
The agent's request. Please check the east-west GPU fabric for Altair AI Lab: 32 GPU nodes, 256 GPUs in all, on N9164E-NS4-O leaves running NX-OS 10.6(1)F and 2 N9164E-NS4-O spines running NX-OS 10.6(1)F, linked with OSFP-800G-DR8 optics over 200 metres, with Spectrum-X and the DCN Advantage license, PFC and ECN lossless, stated as 1.0 to 1. It is ready for the build team.

    
Verdict: Allowed.

    
6. Release, license and stated figures: DCN Essentials, spines on 10.5(3)F, wrong GPU count

    
The agent's request. Please check the east-west GPU fabric for Cobalt Cloud: 64 GPU nodes, 1,024 GPUs in all, on N9364E-SG2-O leaves running NX-OS 10.6(1)F and 4 N9364E-SG2-O spines running NX-OS 10.5(3)F, linked with OSFP-800G-DR8 optics over 250 metres, with Spectrum-X and the DCN Essentials license, PFC and ECN lossless, stated as 1.0 to 1. It is ready for the build team.

    
Verdict: Blocked.

    
    
      
  • Spectrum-X needs a newer NX-OS on the spine. Spectrum-X needs NX-OS 10.6(1)F or later (Cisco data sheet); spine N9364E-SG2-O is set to run 10.5(3)F.
    • 
      
  • Spectrum-X needs the DCN Advantage license. Spectrum-X needs the DCN Advantage license (Cisco data sheet); the proposal includes DCN Essentials.
    • 
      
  • Stated GPU count does not match. Stated GPU count does not match: the agent stated 1024; the validator computes 512 (64 nodes x 8 GPUs).
    • 
    

    
7. Release, license and stated figures: DCN Advantage, 10.6(1)F, correct count

    
The agent's request. Please check the east-west GPU fabric for Cobalt Cloud: 64 GPU nodes, 512 GPUs in all, on N9364E-SG2-O leaves running NX-OS 10.6(1)F and 4 N9364E-SG2-O spines running NX-OS 10.6(1)F, linked with OSFP-800G-DR8 optics over 250 metres, with Spectrum-X and the DCN Advantage license, PFC and ECN lossless, stated as 1.0 to 1. It is ready for the build team.

    
Verdict: Allowed.

    
8. Beyond the reference design: 144 nodes, ECMP, no lossless

    
The agent's request. Please check the east-west GPU fabric for Orchid Systems: 144 GPU nodes, 1,152 GPUs in all, on N9364E-SG2-O leaves running NX-OS 10.5(3)F and 9 N9364E-SG2-O spines running NX-OS 10.5(3)F, linked with OSFP-800G-DR8 optics over 300 metres, with Plain ECMP and the DCN Advantage license, no lossless configuration, stated as 1.0 to 1. It is ready for the build team.

    
Verdict: Needs approval.

    
    
      
  • Larger than the reference architecture. 144 nodes is above the 128 nodes (1,024 GPUs) of the Cisco reference architecture; confirm the design with Cisco.
    • 
      
  • Plain ECMP load balancing. Plain ECMP proposed instead of Spectrum-X with fine-grain load balancing; Cisco positions Spectrum-X for AI east-west fabrics.
    • 
      
  • No lossless configuration. No lossless configuration proposed; RoCE east-west fabrics are normally run lossless with PFC and ECN.
    • 
    

  

  

    
Narration transcript
Click to expand
    
The text of the narrated walkthrough, step by step.

    
    
      
  1. Welcome to a demo of Predictika's Validator. An SE, a partner or an AI agent proposes the east-west fabric for a GPU cluster: which Cisco Nexus 9000 leaf and spine switches, how many spines, which optic, and more. Before the design goes to a simulator, a lab or an order, the proposal goes to the Predictika validator. It recomputes the GPU count, the leaf count and the oversubscription from the nodes and spines, checks every optic against the ports and the distance, checks the Spectrum-X release and license, compares the figures the proposal states with the computed ones, and returns every reason in one answer. Hard rules block; soft recommendations route to a person for approval. This checks that a design is legal, not how it performs: it does not simulate traffic. Nothing here is priced. This is an illustrative model built from Cisco's public documentation. Cisco Systems is not involved with it and has not endorsed it.
    2. 
      
  2. I will start with the first request by the agent.
    3. 
      
  3. The agent says. Please check the east-west GPU fabric for Sable Ridge AI Cloud: 128 GPU nodes, 1,024 GPUs in all, on N9364E-SG2-O leaves running NX-OS 10.5(3)F and 3 N9364E-SG2-O spines running NX-OS 10.6(1)F. It is ready for the build team. The rest of the request is on screen.
    4. 
      
  4. Before it goes out, the agent calls the validator.
    5. 
      
  5. Blocked, with five findings, all in one answer. The fabric is 2.67 to 1 oversubscribed: 1024 leaf-to-spine uplinks are needed for 1:1 but 3 x N9364E-SG2-O give 384 ports. 1:1 needs 8 spines. Optic QDD-400G-DR4 has the QSFP-DD form factor, but leaf N9364E-SG2-O has OSFP ports and spine N9364E-SG2-O has OSFP ports; the optic must fit both. The leaf-to-spine run is 700 m but QDD-400G-DR4 reaches 500 m over single-mode fiber. The other two are on screen.
    6. 
      
  6. The validator returned five blocking findings in one answer. A check that reads one field at a time would pass most of this; it shows only when the choices in the design are checked against each other.
    7. 
      
  7. What this shows: for optics against distance and form factor, a rules-based application is error-prone for this.
    8. 
      
  8. Now the agent resubmits.
    9. 
      
  9. The corrected request. It changes the Leaf NX-OS release to 10.6(1)F and the leaf-to-spine optic to OSFP-800G-DR8.
    10. 
      
  10. It calls the validator again.
    11. 
      
  11. Still blocked, with three findings, all in one answer. The fabric is 2.67 to 1 oversubscribed: 1024 leaf-to-spine uplinks are needed for 1:1 but 3 x N9364E-SG2-O give 384 ports. 1:1 needs 8 spines. The leaf-to-spine run is 700 m but OSFP-800G-DR8 reaches 500 m over single-mode fiber. Stated oversubscription does not match: the agent stated 1.0 to 1; the validator computes 2.67 to 1.
    12. 
      
  12. Two findings cleared. Still blocking: fabric is oversubscribed, optic cannot reach the distance and stated oversubscription does not match.
    13. 
      
  13. The agent resubmits once more.
    14. 
      
  14. The revised request. It changes the spines to 8 and the leaf-to-spine distance to 400 m.
    15. 
      
  15. It calls the validator once more.
    16. 
      
  16. This time the validator clears it.
    17. 
      
  17. The validator also returns the line-item quote. It starts with 8-GPU nodes and N9364E-SG2-O leaf switches; the other two lines are on screen.
    18. 
      
  18. All three earlier findings are cleared, and the validator's figure matches the design.
    19. 
      
  19. Five problems caught in one call, fixed in two passes, and a clean design. That is the value of Predictika's validator: catching mistakes before they reach the build. Now you can try the other scenarios yourself.
    20.