Mirantis validates k0rdent AI on Supermicro servers with AMD MI325X for AI GPU clusters

Mirantis Validates k0rdent AI on Supermicro Bare-Metal Hardware with AMD Instinct MI325X Accelerators

Mirantis has announced the validation of its k0rdent AI platform on Supermicro’s bare-metal infrastructure, featuring AMD Instinct MI325X accelerators. This validation encompasses the entire GPU cluster supply chain, from the automated provisioning of physical servers to the execution of large language model (LLM) inference tasks—an architecture Mirantis refers to as Metal-to-Model.

Deploying a GPU cluster for enterprise AI involves navigating two intertwined challenges: creating an efficient processing chain and bringing it online quickly without configuration complexities. Mirantis details these aspects in its validation report, which outlines the integration of k0rdent AI on Supermicro bare-metal servers outfitted with AMD Instinct MI325X accelerators.

The validation integrates offerings from three key players: Mirantis for the control plane and orchestration, Supermicro for the bare-metal servers, and AMD for the GPU accelerators and high-performance network cards. The testbed comprised two GPU nodes (AS-8126GS-TNMR), each featuring two AMD EPYC 9965 processors with 192 cores and eight AMD MI325X accelerators, plus a controller node (AS-2124BT-HNTR) and a k0rdent management node, all interconnected via AMD Pensando Pollara 400 network cards operating at 400 GbE. The virtualization layer utilized KubeVirt HCO v1.7.0, while the control plane ran on k0rdent Enterprise v1.2.2.

The implications of the report extend beyond hardware validation, addressing the needs of organizations deploying GPU clusters for sovereign AI or hybrid GPU clouds that aim to minimize latency between hardware delivery and workload production. k0rdent AI automates this continuum, managing everything from BIOS configuration to GPU operator deployment.

Automated Bare-Metal Provisioning

The k0rdent Bare Metal Operator, built on the open-source Metal3 and Ironic projects, manages the full lifecycle of Supermicro’s physical nodes via the Redfish protocol connected to the BMC controller. The validation illustrates complete automation of BIOS configuration, firmware updates, and RAID orchestration without manual intervention. Supermicro nodes appear as BareMetalHost objects in the Kubernetes API, allowing them to be treated declaratively alongside any other cloud-native resource.

The PXE-less automated boot process deploys a hardened operating system and the k0s Kubernetes distribution across the entire fleet. This approach removes the overhead traditionally associated with hypervisors between the hardware and Kubernetes, ensuring a reproducible deployment model compliant with Cluster API (CAPI) standards. For teams managing GPU server farms for AI, this reproducibility significantly reduces the risk of inconsistencies between environments and minimizes recovery times after incidents.

Additionally, the validation confirms compatibility between the k0rdent control plane and Supermicro’s modular architecture, crucial for organizations looking to deploy GPU clusters at scale without relying on an hyperscaler—especially important in sovereign AI contexts where data locality and infrastructure control are regulatory constraints.

Automated Deployment of AMD GPU and Network Operators

The AMD GPU Operator is deployed through the k0rdent catalog as a ready-to-use ServiceTemplate. It automatically integrates the ROCm 7.0 stack and configures the container engine for GPU resource access. The validation employs the ROCm Validation Suite to confirm peer-to-peer bandwidth and memory throughput among the eight interconnected GPUs leveraging AMD’s Infinity Fabric, a key architectural feature for high-performance distributed computing in large-scale training and inference workloads.

Meanwhile, the AMD Network Operator v1.0.0 automates driver installation for AMD Pensando Pollara 400 network cards with 400 GbE, facilitating high-speed GPU-to-GPU communication within the cluster. The presence of dedicated 400 GbE network cards is essential for maintaining efficient inter-GPU exchanges in high-density configurations—a frequent bottleneck in poorly dimensioned GPU clusters.

The harmonious deployment of both GPU and network operators from the same k0rdent catalog ensures consistent configuration across the fleet, mitigating the need for ad-hoc installation scripts or configuration drift between nodes. For MLOps architects, this uniformity minimizes diagnostic challenges during performance issues in production.

Achieving 160.6 TFLOP in FP64 Tensor Performance

GPU performance benchmarks were conducted following AMD’s official methodology, utilizing an optimized ROCm vLLM container on a single GPU setup. The selected test model was amd/Llama-3.1-8B-Instruct-FP8-KV, executed through a custom PyTorch script measuring raw compute throughput in FP64 tensor and vector precision.

The results showcased an impressive 160.6 TFLOP in FP64 Tensor/Matrix, compared to an AMD reference of 163.4 TFLOP, resulting in a mere 1.7% difference. In FP64 Vector, the benchmarks hit 82.4 TFLOP, exceeding the AMD reference of 81.7 TFLOP—slightly above the manufacturer’s specifications. These figures validate the performance integrity maintained by the k0rdent orchestration layer and confirm that the hardware operates within AMD’s specified ranges for this accelerator model.

For CIOs exploring alternatives to Nvidia GPUs for their sovereign AI infrastructures, these findings provide a measurable benchmark on certified production configuration hardware. The availability of published performance data across a complete stack—Supermicro hardware, AMD accelerators, and k0rdent orchestration—reduces technical uncertainty in GPU infrastructure projects outside of hyperscaler and Nvidia ecosystems.

Scroll to Top