Supercharging Supermicro Bare Metal with Mirantis k0rdent AI for Sovereign AI and Hybrid GPU Clouds

In the high-stakes arena of sovereign AI and hybrid GPU clouds, organizations are choosing full-stack AI infrastructure solutions to accelerate GPU operationalization, ensure efficient utilization, and enforce security and compliance at scale . Mirantis k0rdent AIis a turnkey, production-ready "super control plane" for managing these complex environments, that automates provisioning, lifecycle management, and orchestration of infrastructure and core services, from Metal-to-Model™.

This report details the successful validation of Supermicro’s modular server architecture with the k0rdent AI ecosystem. For this verification, the models GPU A+ ServerAS-8126GS-TNMR and BigTwin A+ Server AS-2124BT-HNTR were used. By unifying bare-metal provisioning, AMD-powered acceleration, and modern virtualization, this stack provides a blueprint for the next generation of AI data centers.

1. The Foundation: k0rdent Baremetal Operator

Validation begins at the physical layer. The k0rdent Baremetal Operator (utilizing Metal3 and Ironic) serves as the bridge between the declarative Kubernetes API and Supermicro hardware.

• Bare Metal Integration: k0rdent communicates directly with the Supermicro Baseboard Management Controller (BMC) through the Redfish protocol to configure all nodes from the metal up. This includes automated BIOS configuration, firmware updates, and RAID orchestration without human intervention.

• Provisioning Workflow: During validation, Supermicro nodes were discovered as BareMetalHost objects. k0rdent successfully automated the PXE-less booting process, deploying a hardened host OS and the k0s Kubernetes distribution across the fleet.

2. AI at Scale: AMD GPU & Network Operators

To satisfy the demands of Generative AI, the validation environment utilized AMD Instinct and MI325X accelerators. The integration was managed via two critical operators:

AMD GPU Acceleration

The AMD GPU Operator was deployed via the k0rdent catalog, which provides ready-to-deploy ServiceTemplates.

• Automated ROCm™ Stack: The operator automatically injected the ROCm 7.0+ driver stack and configured the container runtime.

• Validation Result: Utilizing the ROCm Validation Suite (rvs), the system confirmed peer-to-peer (P2P) bandwidth and memory throughput, ensuring the 8-GPU "fully-meshed" Infinity Fabric™ was operating at peak efficiency.

AMD Network Operator

High-performance networking enables efficient GPU-to-GPU communication in an AI cluster. The AMD Network Operator (v1.0.0) was validated for its ability to:

• Configure AI NICs:  Setup of AMD Pensando™ Pollara 400 NICs.

4. Hardware Test Bench

5. Validation Results Summary

The validation process concluded with a suite of tests to ensure "Day 1" production readiness:

• Provisioning: Bare metal provisioning with k0rdent

  • The validation successfully demonstrated the automated provisioning of production-grade Kubernetes clusters on Supermicro bare-metal hardware using k0rdent’s declarative orchestration engine and the Bare Metal Operator (BMO). 

  • By leveraging the integration of Metal3 and Ironic, k0rdent managed the entire lifecycle of the Supermicro nodes—from out-of-band discovery via BMC/IPMI and hardware introspection to automated OS imaging and Kubernetes bootstrapping. This process eliminated manual configuration and hypervisor overhead, providing a high-performance, consistent, and repeatable deployment model that adheres to Cluster API (CAPI) standards. 

  • The validation confirms that k0rdent effectively bridges the gap between physical server management and cloud-native agility, making it an ideal solution for resource-intensive workloads requiring direct hardware access and deterministic performance on Supermicro infrastructure.

• GPU Validation:

  • To get our cluster up and running for high-performance AI workloads, we used k0rdent to configure the environment with the ROCm and amdgpu-dkms packages to ensure full hardware compatibility. 

  • On the orchestration side, we deployed Kubernetes with k0rdent utilizing Cert-manager 1.15.1 for certificate management and the AMD GPU Operator 1.4.1 to streamline driver and resource provisioning across the fleet.

  • To wrap things up, we confirmed that all 8 GPUs per node were correctly registered and available within the cluster. This was verified using custom scripts built atop the official rocm/vllm container image, leveraging a suite of tools including rocm-smi, Python, and PyTorch to ensure the hardware wasn't just visible, but fully functional for AI models.

• GPU Performance Evaluation: 

  • We conducted performance benchmarks following theofficial AMD documentation. The evaluation environment utilized the rocm/vllm Docker image, focusing on a single GPU configuration as per the official testing framework.

  • Test Configuration

    • Model:amd/Llama-3.1-8B-Instruct-FP8-KV

    • Environment:ROCm-optimized vLLM container

    • Scope: Single GPU unit performance

  • Validation Methodology

    • To validate the performance metrics, we utilized a custom PyTorch script to measure raw compute throughput across different precisions. Specifically, we evaluated the TFLOPs for the following:

Final Word

Validating Supermicro hardware with Mirantis k0rdent AI represents a shift from "building" clusters to "composing" them. 

By leveraging the automated operators for AMD hardware and the flexibility of k0rdent AI and KubeVirt, enterprise customers can run their entire portfolio—from legacy apps to cutting-edge LLMs—on a single, unified, bare-metal platform with automatic deployment and comprehensive platform management from the bare metal up.

The shift from manual to automated deployment and management reduces the workload on the cluster management teams and eliminates human error and inconsistencies from the deployment and management process.

To learn more about deploying Mirantis k0rdent AI on Supermicro GPU servers, contact us.