IBM Maximo on Bare Metal
IBM Maximo Application Suite on Red Hat OpenShift installed directly on physical servers, with NVIDIA GPU worker nodes for Maximo Visual Inspection and Single Node OpenShift for small or disconnected sites
Overview
Maximo Application Suite runs on Red Hat OpenShift, and OpenShift can be installed directly on physical servers. Red Hat documents several bare metal paths: the Assisted Installer, a wizard in OpenShift Cluster Manager; the Agent-based Installer, which builds a bootable image and suits disconnected sites; and installer-provisioned or user-provisioned infrastructure for full control. Bare metal makes the most sense when you own the hardware already, when Maximo Visual Inspection needs NVIDIA GPUs without a passthrough layer, or when a plant or remote site needs a small cluster close to the equipment. The trade-off is that your team owns firmware, disk replacement, and node recovery that a cloud or virtualization platform would otherwise handle. We plan the topology, size the nodes against IBM's requirements, install OpenShift and its storage, set up GPU nodes, and deploy MAS with the IBM MAS CLI.
What teams weigh before running Maximo on Bare Metal
Bare metal details that catch MAS projects
Maximo Visual Inspection needs a ReadWriteMany storage class, so the LVM Storage used on many single-node clusters is not enough on its own for MVI. NVIDIA Pascal GPUs, including the P100, are not supported in Maximo Visual Inspection 9.2 and later. IBM limits Single Node OpenShift to MAS with Maximo Manage for up to 70 concurrent users. Maximo Visual Inspection Edge is installed on a standalone Linux host with Podman or Docker, not on OpenShift.
Reference architecture
Three control plane nodes and at least two workers for production, installed with the Assisted Installer, the Agent-based Installer, or installer-provisioned infrastructure that uses the servers' baseboard management controllers. A compact three-node cluster or Single Node OpenShift fits smaller sites.
OpenShift Data Foundation on local SSD or NVMe drives through the Local Storage Operator provides block storage, a CephFS file system with ReadWriteMany access, and object storage. Single-node clusters typically use LVM Storage on a dedicated disk, with NFS added where shared file storage is needed.
Servers with NVIDIA GPUs join the cluster as workers. Red Hat's Node Feature Discovery Operator labels the hardware, and the NVIDIA GPU Operator installs the driver, container toolkit, device plugin, and DCGM monitoring so Visual Inspection pods can request GPUs.
The IBM MAS CLI installs MAS core, Maximo Manage, and other applications along with dependencies such as MongoDB, Db2, and IBM Suite License Service through Tekton pipelines on the cluster.
Maximo Visual Inspection Edge runs on a separate x86 Linux server with an NVIDIA GPU at the inspection point, using models trained on the central MVI instance.
Trade-offs
70 concurrent users on Single Node OpenShift. IBM's limit for MAS with Maximo Manage on a single-node cluster
16 GB gPU memory for MVI training. Minimum GPU memory IBM requires during Maximo Visual Inspection model training
10,000 random read/write IOPS for MVI. Minimum storage performance IBM lists for Visual Inspection data set uploads and downloads
Assisted, Agent-based, IPI, and UPI installs differ in how much automation they provide and whether they need internet access during install. The wrong choice makes later node additions and upgrades harder.
Manage attachments and Visual Inspection data sets need ReadWriteMany storage, Db2 and MongoDB want fast block storage, and MVI needs at least 10,000 IOPS for data set work. Local disks have to be laid out with all of that in mind.
The NVIDIA GPU Operator, the OpenShift version, the GPU model, and the MVI release all have to line up. Older GPUs such as Pascal cards lose MVI support in 9.2.
Plant networks often block direct internet access. OpenShift, the operators, and the MAS images all have to come from a mirrored registry, and that mirror has to be kept current for every upgrade.
There is no cloud provider to replace a failed node. Firmware updates, disk failures, and node rebuilds become part of your operations runbook.
What we do
We choose the installer that fits your network and team, define control plane, worker, storage, and GPU node roles, and install OpenShift on your servers. For disconnected sites we set up the mirror registry before install.
AI on Bare Metal
GPU nodes can be tainted so that only Visual Inspection and other GPU workloads schedule on them, keeping Manage and core services on general workers.
We deploy OpenShift Data Foundation on local SSD or NVMe drives for block, ReadWriteMany file, and object storage, or LVM Storage plus NFS for single-node clusters, and map each MAS application to the right storage class.
AI on Bare Metal
Visual Inspection storage is sized from IBM's guidance on data set volume and model count, with headroom for growth.
We install Red Hat's Node Feature Discovery Operator and the NVIDIA GPU Operator, validate GPU access with CUDA test containers, and then deploy Visual Inspection. IBM requires GPUs to train all MVI model types and at least 16 GB of GPU memory during training.
AI on Bare Metal
Trained models can be deployed on the cluster for GPU or CPU inference, or exported to Maximo Visual Inspection Edge hosts at the production line.
IBM supports Single Node OpenShift for MAS with Maximo Manage for up to 70 concurrent users, including satellite or disconnected deployments and edge sites where high availability is not critical. We install it on one physical server with LVM Storage.
AI on Bare Metal
Where a site needs camera-based inspection, MVI Edge on its own GPU server sits beside the SNO node and sends results back to the central MAS instance.
We document node replacement, certificate rotation, OpenShift and MAS upgrade sequencing, and GPU Operator upgrades so your team can run the platform.
AI on Bare Metal
DCGM metrics from the GPU Operator show GPU use and health in OpenShift monitoring, which helps decide when to add GPU capacity.
Walk through your cluster, databases, integrations, and AI plans with an engineer who installs and runs MAS on OpenShift.
The stack
Requirements
Red Hat's guide to installer-provisioned and user-provisioned OpenShift installs on physical servers.
Preparing an on-premises install with a bootable agent image, including disconnected installs and single-node or compact topologies.
Worker node memory, ReadWriteMany storage, capacity sizing, and IOPS for Visual Inspection on MAS.
Certified NVIDIA GPU families, GPU memory for training, CPU inference, and the NFD and NVIDIA GPU Operators.
NVIDIA's documentation for the operator that manages GPU drivers, the container toolkit, the device plugin, and DCGM monitoring.
Deploying ODF with the Local Storage Operator on local disks of bare metal nodes.
Operating systems, NVIDIA GPU and driver, memory, and container runtime requirements for MVI Edge hosts.
Yes. MAS runs on Red Hat OpenShift, and OpenShift can be installed directly on physical x86 servers. IBM's Single Node OpenShift guidance covers bare metal installs, and full multi-node clusters can be built with Red Hat's bare metal installers.
The Assisted Installer is a guided wizard in Red Hat's OpenShift Cluster Manager and needs internet access. The Agent-based Installer builds a bootable image locally and is the usual choice for disconnected sites. Installer-provisioned infrastructure automates provisioning through the servers' baseboard management controllers, and user-provisioned infrastructure gives full manual control.
Bare metal fits when you already own suitable servers, when GPU nodes for Visual Inspection would otherwise need passthrough configuration, or when a site needs a small cluster on one or three physical machines. Virtualization is usually easier when you already run a mature VMware or OpenShift Virtualization estate and want live migration and snapshot-based recovery.
IBM states that Visual Inspection requires GPUs to train all model types and needs at least 16 GB of GPU memory during training. In some configurations, trained models can run inference on CPU only, or on an MVI Edge device.
IBM's GPU FAQ lists NVIDIA Ampere, Turing, and Volta families such as the T4, V100, A10, and A100 as certified. Pascal GPUs, including the P100, are not supported in MVI 9.2 and later. Check IBM's compatibility reports for your exact MVI and OpenShift versions.
Install Red Hat's Node Feature Discovery Operator so GPU hardware is detected, then the NVIDIA GPU Operator, which installs the driver, container toolkit, and device plugin. IBM recommends the official Red Hat NFD Operator rather than the community version. Validate with a CUDA test container before installing Visual Inspection.
OpenShift Data Foundation on local SSD or NVMe disks through the Local Storage Operator is the common choice, because it provides block, ReadWriteMany file, and object storage. Single Node OpenShift usually uses LVM Storage on a dedicated disk. Visual Inspection requires ReadWriteMany storage, so LVM Storage alone is not enough for MVI.
IBM says a single-node cluster can be used in production when redundancy, scalability, and high availability are not critical. It is limited to MAS with Maximo Manage for up to 70 concurrent users and also fits satellite or disconnected sites, upgrades of small Maximo deployments to MAS, and proofs of concept.
No. MVI Edge is installed on a standalone Linux server running Red Hat Enterprise Linux 9 with Podman or Ubuntu with Docker. IBM requires an NVIDIA GPU and at least 64 GB of RAM for a local deployment, and the Edge host runs models trained in the central Visual Inspection instance.
Yes. OpenShift supports disconnected installs, typically with the Agent-based Installer and a mirror registry that holds OpenShift, operator, and IBM MAS images. The NVIDIA GPU Operator also has a documented disconnected install path, and every mirrored image has to be refreshed before each upgrade.
Tell us where Maximo runs today and where you want it to run. We will tell you honestly whether Bare Metal is the right fit.
