Predictika · GPU server and AI cluster quotes

An AI sales agent's Relion GPU server quote, checked before it goes out

An AI sales agent prepares a quote for Relion GPU servers: a model, GPU type and count, CPU generation, memory, drives, compute NICs, node count, rack type with its power and space, ClusterWare, a support term and a discount. Before it goes out, the agent sends it to Predictika, which checks it against a declarative model of the product line and returns a verdict and every reason.

Scope: Relion GPU servers (XE2318GT, XE4318GT, XE4418GT, XO1314GT, XE8318GTS and XE8418GTS) with GPU type and count, CPU generation, memory, drives, compute NICs, rack power and space, ClusterWare, support term and discount. See everything this validator covers

This validator models Relion GPU server quotes, not just the scenarios shown below.

  • Model, GPU and CPU — which GPU types each Relion model takes, how many GPUs it holds, the HGX baseboards that are sold with all eight GPUs, and the CPU generation each model uses.
  • Memory, drives and networking — the memory maximum and the memory the GPUs need, the drive bays, and the compute NIC count for HGX and PCIe-GPU nodes.
  • Rack, cluster, support and discount — rack type, rack power and space for the nodes, ClusterWare for larger clusters, the support term, discount limits and the approvals for large orders, big discounts and B300 allocation.

Any Relion GPU server quote that touches these rules is in scope — not only the scenarios pictured here. Not covered: other Penguin Solutions products and services. Model names, form factors, GPU slot counts, CPU generations, the memory maximum and drive bays of the XE2318GT and XE4318GT, and the 3-year warranty with up to 4 years of on-site support follow public Penguin Solutions pages; every other rule, the approval thresholds and all prices are modeled and are not published by Penguin Solutions (see the note at the foot of the page).

The AI sales agentBuilds the GPU server quote from the conversation and says it is ready.
Predictika validatorChecks every rule at once and returns the verdict, each finding, and the computed figures.

6 scenarios in 3 groups · each group runs from the agent’s first attempt to the outcome · click one to switch

What the agent proposes

The AI sales agent says

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

Proposing agent: AI sales 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
GPU, CPU and model pairingsOnly with care
Which GPU types and CPU generation each model takes differs model by model, so every rule is keyed to the model and several other fields at once.		The model, GPU type, CPU generation and counts are fields of one quote and are checked together on every call.
Rack power and space for the nodesOnly with workarounds
The nodes drawn against the rack supply and the rack units plus switches have to be multiplied and compared across several fields.		Node count, node power, rack units and the rack supply are computed and checked together on every call.
Memory the GPUs needOnly with workarounds
Memory has to be compared with a floor that depends on the GPU count and with a ceiling that depends on the model.		The floor and the ceiling are computed from the quote and compared on every call.
Every problem returned in one answerOnly with workarounds
Rules tend to fire as the quote is saved or a step is completed, so a rep often 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 quote; approval findings need a person to sign off; everything else is computed.

    
1The selected CPU generation does not match the selected Relion model.
2The selected GPU type is not offered on the selected Relion model.
3The XO series is built for 21-inch OCP racks and the XE series for 19-inch racks; the rack type does not match the model.
45 GPUs per node exceeds the 4 the model holds.
5HGX models are quoted with the full 8-GPU baseboard; the quote has 4 GPUs per node.
64 compute NICs per node does not fit this model; HGX models take 8 and PCIe-GPU models 0 to 4.
76 compute NICs per node does not fit this model; HGX models take 8 and PCIe-GPU models 0 to 4.
8HGX B300 systems are allocation-limited; confirm availability and lead time.
98,192 GB of memory per node exceeds the 6,144 GB maximum for this model.
10512 GB of memory per node is below the 1,024 GB needed for the GPUs (128 GB per GPU).
1113 drives per node exceeds the 12 drive bays of this model.
129 GPUs per node exceeds the 8 the model holds.
13A GPU server quote needs at least one GPU per node; the quote has 0.
14The nodes draw 43,200 W and the rack supplies 40,000 W.
15The nodes need 40 rack units plus 4 for switches, and the rack has 42.
16The standard warranty is 3 years with up to 4 years of on-site support; the quote has 5 years.
17A cluster of 8 nodes is quoted with Penguin ClusterWare software.
18An order of 16 nodes needs a sales manager review.
19A discount of 15 percent needs sales manager approval.
20A discount of 30 percent is above the 25 percent limit.
21The quote has 0 nodes; at least one is needed.
2210 drives per node exceeds the 8 drive bays of this model.

  

  

    
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. Model, GPU, CPU, memory and drives: An XE4418GT with an unsupported GPU type, the wrong CPU generation, too little memory and too many drives

    
The agent's request. Please quote 4 XE4418GT nodes with 8 H100 PCIe GPUs each, 4th/5th Gen Xeon, 512 GB of memory, 14 drives and 4 compute NICs per node, in a 19-inch rack supplying 40,000 W and 42 rack units, without ClusterWare, 3 years of support and a discount of 5 percent. It is ready to send.

    
Verdict: Blocked.

    
    
      
  • GPU type not offered on this model. The selected GPU type is not offered on the selected Relion model.
    • 
      
  • CPU generation does not match model. The selected CPU generation does not match the selected Relion model.
    • 
      
  • Memory below 128 GB per GPU. 512 GB of memory per node is below the 1,024 GB needed for the GPUs (128 GB per GPU).
    • 
      
  • More drives than bays. 14 drives per node exceeds the 12 drive bays of this model.
    • 
    

    
2. Model, GPU, CPU, memory and drives: The agent resubmits after fixing the errors

    
The agent's request. Please quote 4 XE4418GT nodes with 8 H200 NVL GPUs each, Xeon 6, 2,048 GB of memory, 8 drives and 4 compute NICs per node, in a 19-inch rack supplying 40,000 W and 42 rack units, without ClusterWare, 3 years of support and a discount of 5 percent. It is ready to send.

    
Verdict: Allowed.

    
    
      
  • Quote summary. 4 node(s), 32 GPUs in total, 30,400 W and 16 rack units; illustrative total 1,220,681 USD after discount.
    • 
    

    
3. HGX GPUs, NICs, rack power and space: Five HGX B200 nodes with half the GPUs, too few NICs and a rack that cannot power or hold them

    
The agent's request. Please quote 5 XE8318GTS nodes with 4 HGX B200 GPUs each, 4th/5th Gen Xeon, 2,048 GB of memory, 8 drives and 4 compute NICs per node, in a 19-inch rack supplying 40,000 W and 42 rack units, without ClusterWare, 3 years of support and a discount of 5 percent. It is ready to send.

    
Verdict: Blocked.

    
    
      
  • HGX models are sold with all 8 GPUs. HGX models are quoted with the full 8-GPU baseboard; the quote has 4 GPUs per node.
    • 
      
  • Compute NIC count does not fit the model. 4 compute NICs per node does not fit this model; HGX models take 8 and PCIe-GPU models 0 to 4.
    • 
      
  • Rack power exceeded. The nodes draw 71,500 W and the rack supplies 40,000 W.
    • 
      
  • Rack space exceeded. The nodes need 40 rack units plus 4 for switches, and the rack has 42.
    • 
    

    
4. HGX GPUs, NICs, rack power and space: The agent resubmits after fixing the errors

    
The agent's request. Please quote 2 XE8318GTS nodes with 8 HGX B200 GPUs each, 4th/5th Gen Xeon, 2,048 GB of memory, 8 drives and 8 compute NICs per node, in a 19-inch rack supplying 40,000 W and 42 rack units, without ClusterWare, 3 years of support and a discount of 5 percent. It is ready to send.

    
Verdict: Allowed.

    
    
      
  • Quote summary. 2 node(s), 16 GPUs in total, 28,600 W and 16 rack units; illustrative total 741,820 USD after discount.
    • 
    

    
5. Cluster software, support term and discount: An eight-node cluster without ClusterWare, a 5-year support term and a 30 percent discount

    
The agent's request. Please quote 8 XE4318GT nodes with 8 H100 PCIe GPUs each, 4th/5th Gen Xeon, 2,048 GB of memory, 4 drives and 4 compute NICs per node, in a 19-inch rack supplying 80,000 W and 80 rack units, without ClusterWare, 5 years of support and a discount of 30 percent. It is ready to send.

    
Verdict: Blocked.

    
    
      
  • Support term not offered. The standard warranty is 3 years with up to 4 years of on-site support; the quote has 5 years.
    • 
      
  • Clusters of 8 or more nodes need ClusterWare. A cluster of 8 nodes is quoted with Penguin ClusterWare software.
    • 
      
  • Discount above 25 percent. A discount of 30 percent is above the 25 percent limit.
    • 
    

    
6. Cluster software, support term and discount: The agent resubmits after fixing the errors

    
The agent's request. Please quote 8 XE4318GT nodes with 8 H100 PCIe GPUs each, 4th/5th Gen Xeon, 2,048 GB of memory, 4 drives and 4 compute NICs per node, in a 19-inch rack supplying 80,000 W and 80 rack units, with ClusterWare, 4 years of support and a discount of 8 percent. It is ready to send.

    
Verdict: Allowed.

    
    
      
  • Quote summary. 8 node(s), 64 GPUs in total, 57,600 W and 32 rack units; illustrative total 2,130,179 USD after discount.
    • 
    

  

  

    
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 sales agent prepares a quote for Relion GPU servers: a model, GPU type and count, CPU generation, and more. Before it goes out, the agent sends it to Predictika, which checks it against a declarative model of the product line and returns a verdict and every reason.
    2. 
      
  2. I will start with the first request by the agent.
    3. 
      
  3. The agent says. Please quote 4 XE4418GT nodes with 8 H100 PCIe GPUs each, 4th/5th Gen Xeon, 512 GB of memory, 14 drives and 4 compute NICs per node. It is ready to send. The rest of the request is on screen.
    4. 
      
  4. Before it goes out, the agent calls the validator.
    5. 
      
  5. Blocked, with four findings, all in one answer. The selected GPU type is not offered on the selected Relion model. The selected CPU generation does not match the selected Relion model. 512 GB of memory per node is below the 1,024 GB needed for the GPUs (128 GB per GPU). The other one is on screen.
    6. 
      
  6. The validator returned four 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 quote are checked against each other.
    7. 
      
  7. What this shows: for GPU, CPU and model pairings, a rules-based CPQ is error-prone for this.
    8. 
      
  8. Now the agent resubmits.
    9. 
      
  9. The corrected request. It changes the GPU type to H200 NVL, the CPU generation to Xeon 6, the memory per node to 2,048 GB and the drives per node to 8.
    10. 
      
  10. It calls the validator again.
    11. 
      
  11. This time the validator clears it.
    12. 
      
  12. All four earlier findings are cleared.
    13. 
      
  13. Four problems caught in one call, fixed in one pass, and a clean quote. That is the value of Predictika's validator: catching mistakes before they reach the customer. Now you can try the other scenarios yourself.
    14.