Predictika · Cisco Nexus 9000 AI cluster designs

AI Cluster Order Check for Cisco Nexus 9000 AI Designs

An AI agent prepares an order for an AI training cluster built on Cisco's Nexus 9000 reference architecture: UCS C885A M8 nodes with 8 NVIDIA H200 GPUs, BlueField-3 NICs, and Nexus 9000 east-west switches. Before the order is placed, the agent sends it to Predictika, which checks it against a declarative model of the rules in Cisco's public data sheet and returns a verdict, every reason and the figures it computed. The order is released only when the verdict is allow, or after an approval. This validator contains no prices.

Scope: Cisco Nexus 9000 AI cluster designs (N9364E-SG2-O, N9164E-NS4-O and N9332D-GX2B east-west switches, UCS C885A M8 nodes, BlueField-3 NICs). See everything this validator covers

This validator handles any cluster order that touches the rules below — not just the scenarios pictured here.

  • Cluster shape — nodes in scalable units of 4, up to 128 nodes (32 units, 1,024 GPUs); 8 GPUs and 8 east-west NICs per node; a spine-leaf fabric above 16 nodes; leaf ports for 8 400G links per node; reference switch counts for 12 nodes (2 leaf), 32 (4 leaf, 2 spine), 64 (8 leaf, 8 spine) and 128 (16 leaf, 16 spine, 8 superspine), with any difference sent to a person.
  • Optics — OSFP-800G-DR8 on the N9364E-SG2-O and N9164E-NS4-O; QDD-400G-DR4 or QSFP-200G-SR4 on the N9332D-GX2B; QSFP-400G-DR4 on BlueField-3 B3140H NICs.
  • Spectrum-X — NX-OS 10.6(1)F or later, the DCN Advantage license and the FGLB license (FGLB not required for east-west with the N9164E-NS4-O).
  • Storage — at least 8 400G links to the storage network, as the data sheet gives for its storage design; and the agent's stated GPU and scalable-unit totals against the computed ones.

Not covered: prices; the north-south network and NIC count (the data sheet says both one B3220 and two B3240 per node, so it is not modeled); the switch count for 16 nodes (the data sheet's bill of materials lists 4 800G switches while its sizing table lists 2 leaf and no spine); cabling quantities, power and cooling; whether NX-OS 10.6(1)F and DCN Advantage are still needed with the N9164E-NS4-O (the data sheet says only that the Spectrum-X license is not required there). The F train of the NX-OS release is assumed. Cluster design, release numbers and licenses are supplied by the customer's own systems and are not Cisco facts.

The AI ordering agentAssembles the cluster order and says it is ready to place.
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 ordering agent says

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

Proposing agent: AI ordering 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) 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
Cluster shape from published reference designs: nodes in scalable units of 4, at most 128 nodes, a spine-leaf fabric above 16 nodes, and the leaf, spine and superspine counts Cisco lists for 12, 32, 64 and 128 nodesOnly with workarounds
A CPQ knows quantities of parts, not the topology the parts must form. Each reference size becomes its own rule and the arithmetic (nodes times 8 links against leaf ports) has to be written out.		Node, switch and port counts are in one request and are checked against the reference design together.
Which optic fits which port: OSFP-800G-DR8 on the 800G switches, QDD-400G-DR4 or QSFP-200G-SR4 on the N9332D-GX2B, QSFP-400G-DR4 on BlueField-3 NICsOnly with care
Compatibility is a pair rule for every switch and optic. It works until a new switch or optic is added; then every pair rule needs another look.		Switch model and optic are checked together on every call.
Spectrum-X dependencies: NX-OS 10.6(1)F or later, the DCN Advantage license and the FGLB license (not needed for east-west with the N9164E-NS4-O)Only with workarounds
A release comparison plus a license rule with an exception. The release number is usually a free-text field that has to be parsed.		Release, licenses and switch model travel with the request and are compared together.
A GPU or scalable-unit total stated by an outside AI agent compared with the computed valueNot what it is built for
A CPQ computes its own totals; it has no field for a number an agent asserts, so a wrong statement is never compared.		The stated figure is a field of the request and is compared with the computed one on every call.
Where Cisco's own data sheet differs from itself (the north-south NIC count; the 16-node switch count) the validator does not guessOnly when vigilant
A rule either blocks or allows. A documented disagreement needs its own approval path, so it is usually settled by picking one reading.		Those two points are not modeled and the page says so; a design that departs from a listed reference size goes to a person.
Every problem returned in one answerOnly with workarounds
Rules tend to fire as a step is completed, so a requester fixes one error and meets the next.		All findings come back together, whatever order the choices arrived in.

  

  

    
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 request; approval findings need a person to sign off; everything else is computed.

    
122 nodes is not a multiple of 4; the reference design is built from scalable units of 4 nodes.
2132 nodes exceeds the 128 nodes (32 scalable units, 1,024 GPUs) the reference design scales to.
3a cluster of 20 nodes is larger than 16, so the east-west compute fabric must be a spine-leaf design; the request has no spine switches.
4the 2 east-west leaf switches provide 64 400G ports; the 12 nodes need 96 (8 per node).
5the reference node has 8 GPUs; the request has 7 per node.
6the 64-port 800G OSFP switches (N9,364E-SG2-O and N9,164E-NS4-O) take OSFP-800G-DR8 optics; the optic in the request does not fit them.
7the N9,332D-GX2B takes QDD-400G-DR4 (single-mode) or QSFP-200G-SR4 (multimode) optics; the optic in the request does not fit it.
8BlueField-3 B3,140H NICs use QSFP-400G-DR4 optics; the optic in the request does not fit them.
9Spectrum-X on Nexus 9,000 needs NX-OS 10.6(1)F or later; the request has NX-OS 10.6(0).
10Spectrum-X needs the DCN Advantage license; the request does not include it.
11Spectrum-X needs the Fine Grain Load Balancing (FGLB) license in addition to DCN Advantage; the request does not include it.
12the storage design in the data sheet needs at least 8 400G links to the storage network; the request has 6.
13the request states 100 GPUs; 12 nodes with 8 GPUs each give 96.
14the switch counts differ from the reference design the data sheet lists for this cluster size (12 nodes: 2 leaf; 32 nodes: 4 leaf and 2 spine; 64 nodes: 8 leaf and 8 spine; 128 nodes: 16 leaf, 16 spine and 8 superspine); a person should confirm the design.

  

  

    
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. Spectrum-X: A 128-node cluster on older NX-OS without the FGLB license

    
The agent's request. Please order a cluster of 128 Cisco UCS C885A M8 nodes (32 scalable units, 1,024 GPUs), each with 8 GPUs and 8 east-west NICs. The east-west fabric uses 16 leaf, 16 spine and 6 superspine N9364E-SG2-O switches, with OSFP-800G-DR8 optics on the switches and QSFP-400G-DR4 optics on the NICs. Spectrum-X is enabled on NX-OS 10.5(2) with the DCN Advantage license and no FGLB license. It is ready to order.

    
Verdict: Blocked.

    
    
      
  • NX-OS too old for Spectrum-X. Spectrum-X on Nexus 9000 needs NX-OS 10.6(1)F or later; the request has NX-OS 10.5(2).
    • 
      
  • FGLB license missing. Spectrum-X needs the Fine Grain Load Balancing (FGLB) license in addition to DCN Advantage; the request does not include it.
    • 
      
  • Differs from the reference design. The switch counts differ from the reference design the data sheet lists for this cluster size (12 nodes: 2 leaf; 32 nodes: 4 leaf and 2 spine; 64 nodes: 8 leaf and 8 spine; 128 nodes: 16 leaf, 16 spine and 8 superspine); a person should confirm the design.
    • 
    

    
2. Spectrum-X: The agent resubmits after fixing the errors

    
The agent's request. Please order a cluster of 128 Cisco UCS C885A M8 nodes (32 scalable units, 1,024 GPUs), each with 8 GPUs and 8 east-west NICs. The east-west fabric uses 16 leaf, 16 spine and 6 superspine N9364E-SG2-O switches, with OSFP-800G-DR8 optics on the switches and QSFP-400G-DR4 optics on the NICs. Spectrum-X is enabled on NX-OS 10.6(1) with the DCN Advantage license and the FGLB license. It is ready to order.

    
Verdict: Needs approval.

    
    
      
  • Differs from the reference design. The switch counts differ from the reference design the data sheet lists for this cluster size (12 nodes: 2 leaf; 32 nodes: 4 leaf and 2 spine; 64 nodes: 8 leaf and 8 spine; 128 nodes: 16 leaf, 16 spine and 8 superspine); a person should confirm the design.
    • 
      
  • Summary. 128 nodes in 32 scalable units, 1,024 GPUs in total.
    • 
    

    
3. Spectrum-X: The agent resubmits after fixing the errors

    
The agent's request. Please order a cluster of 128 Cisco UCS C885A M8 nodes (32 scalable units, 1,024 GPUs), each with 8 GPUs and 8 east-west NICs. The east-west fabric uses 16 leaf, 16 spine and 8 superspine N9364E-SG2-O switches, with OSFP-800G-DR8 optics on the switches and QSFP-400G-DR4 optics on the NICs. Spectrum-X is enabled on NX-OS 10.6(1) with the DCN Advantage license and the FGLB license. It is ready to order.

    
Verdict: Allowed.

    
    
      
  • Summary. 128 nodes in 32 scalable units, 1,024 GPUs in total.
    • 
    

    
4. Cluster size: 20 nodes with no spine switches

    
The agent's request. Please order a cluster of 20 Cisco UCS C885A M8 nodes (5 scalable units, 160 GPUs), each with 8 GPUs and 8 east-west NICs. The east-west fabric uses 2 leaf and no spine N9364E-SG2-O switches, with OSFP-800G-DR8 optics on the switches and QSFP-400G-DR4 optics on the NICs. It is ready to order.

    
Verdict: Blocked.

    
    
      
  • No spine above 16 nodes. A cluster of 20 nodes is larger than 16, so the east-west compute fabric must be a spine-leaf design; the request has no spine switches.
    • 
    

    
5. Cluster size: The agent resubmits after fixing the errors

    
The agent's request. Please order a cluster of 20 Cisco UCS C885A M8 nodes (5 scalable units, 160 GPUs), each with 8 GPUs and 8 east-west NICs. The east-west fabric uses 4 leaf and 2 spine N9364E-SG2-O switches, with OSFP-800G-DR8 optics on the switches and QSFP-400G-DR4 optics on the NICs. It is ready to order.

    
Verdict: Allowed.

    
    
      
  • Summary. 20 nodes in 5 scalable units, 160 GPUs in total.
    • 
    

    
6. Optics: Optics that do not fit the 800G switches

    
The agent's request. Please order a cluster of 32 Cisco UCS C885A M8 nodes (8 scalable units, 256 GPUs), each with 8 GPUs and 8 east-west NICs. The east-west fabric uses 4 leaf and 2 spine N9364E-SG2-O switches, with QDD-400G-DR4 optics on the switches and QSFP-400G-DR4 optics on the NICs. It is ready to order.

    
Verdict: Blocked.

    
    
      
  • Wrong optic on 800G switches. The 64-port 800G OSFP switches (N9364E-SG2-O and N9164E-NS4-O) take OSFP-800G-DR8 optics; the optic in the request does not fit them.
    • 
    

    
7. Optics: The agent resubmits after fixing the errors

    
The agent's request. Please order a cluster of 32 Cisco UCS C885A M8 nodes (8 scalable units, 256 GPUs), each with 8 GPUs and 8 east-west NICs. The east-west fabric uses 4 leaf and 2 spine N9364E-SG2-O switches, with OSFP-800G-DR8 optics on the switches and QSFP-400G-DR4 optics on the NICs. It is ready to order.

    
Verdict: Allowed.

    
    
      
  • Summary. 32 nodes in 8 scalable units, 256 GPUs in total.
    • 
    

    
8. Node count: A node count outside the reference design

    
The agent's request. Please order a cluster of 130 Cisco UCS C885A M8 nodes (1,040 GPUs), each with 8 GPUs and 8 east-west NICs. The east-west fabric uses 16 leaf, 16 spine and 8 superspine N9364E-SG2-O switches, with OSFP-800G-DR8 optics on the switches and QSFP-400G-DR4 optics on the NICs. It is ready to order.

    
Verdict: Blocked.

    
    
      
  • Not a multiple of 4. 130 nodes is not a multiple of 4; the reference design is built from scalable units of 4 nodes.
    • 
      
  • More than 128 nodes. 130 nodes exceeds the 128 nodes (32 scalable units, 1,024 GPUs) the reference design scales to.
    • 
    

  

  

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

    
    
      
  1. Welcome to a demo of Predictika's Validator. An AI agent prepares an order for an AI training cluster built on Cisco's Nexus 9000 reference architecture: UCS C885A M8 nodes with 8 NVIDIA H200 GPUs, BlueField-3 NICs, Nexus 9000 east-west switches. Before the order is placed, the agent sends it to Predictika, which checks it against a declarative model of the rules in Cisco's public data sheet and returns a verdict, every reason and the figures it computed. The order is released only when the verdict is allow, or after an approval. This validator contains no prices.
    2. 
      
  2. I will start with the first request by the agent.
    3. 
      
  3. The agent says. Please order a cluster of 128 Cisco UCS C885A M8 nodes (32 scalable units, 1,024 GPUs), each with 8 GPUs and 8 east-west NICs. The east-west fabric uses 16 leaf, 16 spine and 6 superspine N9364E-SG2-O switches, with OSFP-800G-DR8 optics on the switches and QSFP-400G-DR4 optics on the NICs. The rest of the request is on screen.
    4. 
      
  4. Before it goes out, the agent calls the validator.
    5. 
      
  5. Blocked, with two findings and one approval, all in one answer. Spectrum-X on Nexus 9000 needs NX-OS 10.6(1)F or later; the request has NX-OS 10.5(2). Spectrum-X needs the Fine Grain Load Balancing (FGLB) license in addition to DCN Advantage; the request does not include it.
    6. 
      
  6. The validator returned two blocking findings and one approval 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 order are checked against each other.
    7. 
      
  7. What this shows: for spectrum-X dependencies: NX-OS 10.6(1)F or later, the DCN Advantage license and the FGLB license (not needed for east-west with the N9164E-NS4-O), a rules-based CPQ can do this only with workarounds.
    8. 
      
  8. Now the agent resubmits.
    9. 
      
  9. The corrected request. It changes the NX-OS release to 10.6(1) and the FGLB license to yes.
    10. 
      
  10. It calls the validator again.
    11. 
      
  11. Not blocked, but held for a person.
    12. 
      
  12. The two blocking findings are cleared; one approval remains for a person: differs from the reference design.
    13. 
      
  13. The agent resubmits once more.
    14. 
      
  14. The revised request. It changes the superspine switches to 8.
    15. 
      
  15. It calls the validator once more.
    16. 
      
  16. This time the validator clears it.
    17. 
      
  17. The earlier finding is cleared, and the validator's figure matches the order.
    18. 
      
  18. Three problems caught in one call, fixed in two passes, and a clean order. That is the value of Predictika's validator: catching mistakes before they reach the customer.
    19.