Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchPC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
Short answer: Run:ai v3.0 supports upstream Kubernetes 1.35–1.37 with containerd, Red Hat OpenShift 4.20–4.22 with CRI-O, and managed EKS, GKE, AKS, OKE, and RKE2 clusters when their Kubernetes version falls inside that upstream range. You also need NVIDIA GPU Operator 25.10–26.7.1, Prometheus, supported Linux nodes, synchronized clocks, cluster-admin access, storage, and prepared ingress, DNS, TLS, load-balancing, and registry connectivity. vGPU, NVIDIA DGX Spark, Jetson, and workstation deployments are not supported.
Run:ai v3.0 compatibility at a glance
Run:ai compatibility is a stack, not a single version check. The Kubernetes or OpenShift release, container runtime, CPU architecture, NVIDIA GPU Operator, driver, monitoring stack, networking, storage, and workload framework all have to pass their respective gates. The following values are the versions listed in NVIDIA’s Run:ai v3.0 support matrix.
| Component | Supported value or condition | What to verify |
|---|---|---|
| Upstream Kubernetes | 1.35–1.37 | Use containerd; x86 and ARM are listed. |
| Red Hat OpenShift | 4.20–4.22 | Use CRI-O; x86 and ARM are listed. |
| Managed Kubernetes | EKS, GKE, AKS, OKE, and RKE2 when based on a supported upstream version | Confirm that the provider actually offers the desired release in your region and account. |
| NVIDIA GPU Operator | 25.10–26.7.1 for Run:ai v3.0 | Check the operator’s data-center GPU support list for the exact model. |
| Monitoring | Prometheus 3.13.1 or Kube-Prometheus Stack 88.6.5 and above | Ensure the monitoring installation is reachable by the Run:ai components. |
| GPU fractions | GPU Operator 26.7.1 or later, CUDA driver r615 or later | Set cdi.nriPluginEnabled=false; NRI must be disabled. |
| Multi-Node NVLink | Kubernetes 1.32 or later and GPU Operator 25.3 or later | Use the additional Network Operator and, where applicable, DRA requirements. |
| Virtualization and edge hardware | vGPU, NVIDIA DGX Spark, NVIDIA Jetson, and workstations unsupported | Use physical GPU passthrough on supported data-center hardware. |
Which Kubernetes and OpenShift versions does Run:ai support?
Upstream Kubernetes
For Run:ai v3.0, the listed upstream range is Kubernetes 1.35 through 1.37, inclusive. The supported container runtime is containerd. Both x86 and ARM node architectures are listed, so a mixed-architecture design can be viable only when every other component, including the GPU Operator and workload images, supports the selected architecture.
Free tools Windows power users keep installed
One-click scans. No signup required.
Do not treat the range as a promise that every patch release has identical behavior. Pin and test the exact patch release you intend to operate, then check the Run:ai matrix again before changing it. A Kubernetes version outside 1.35–1.37 is not covered by this v3.0 statement, even if a neighboring release appears operationally similar.
Red Hat OpenShift
OpenShift 4.20 through 4.22 is listed for Run:ai v3.0, with CRI-O as the container runtime and x86 and ARM architectures supported. OpenShift’s operators, security constraints, and upgrade channels are administered differently from upstream Kubernetes, so validate the exact OpenShift minor release and the corresponding NVIDIA operator packaging rather than copying an upstream installation procedure.
Managed Kubernetes services
EKS, GKE, AKS, OKE, and RKE2 are listed when they track a supported upstream Kubernetes version. This is conditional support: Run:ai’s matrix does not make a provider release available if that provider has not released it for your region, account, control-plane channel, or node image. Check the provider’s own release-availability documentation before scheduling an upgrade.
The same rule applies to provider-specific node behavior. Confirm that you can select the required Linux image, containerd configuration, GPU instance or bare-metal node type, storage integration, load balancer, ingress controller, and network policies. A managed service can satisfy the Kubernetes version requirement while still failing a storage, networking, GPU, or identity prerequisite.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →GPU Operator, drivers, and monitoring requirements
NVIDIA GPU Operator
NVIDIA states that a Run:ai cluster requires NVIDIA GPU Operator installed on the Kubernetes cluster. For Run:ai v3.0, the support matrix lists GPU Operator versions 25.10 through 26.7.1. Select an operator version inside that interval and then check its supported data-center GPU list. There is no single universal minimum GPU model published for Run:ai; model eligibility is delegated to the GPU Operator and the relevant framework support matrices.
Install and validate the operator before applying Run:ai charts or guided installation. The operator must be able to discover the devices, expose them to containers, and report healthy driver and device-plugin state on every intended worker. A control-plane node that has no GPU does not remove the requirement for healthy operator components on GPU workers.
Prometheus and Kube-Prometheus Stack
Monitoring is a baseline prerequisite, not an optional add-on. The v3.0 matrix lists Prometheus 3.13.1 or Kube-Prometheus Stack 88.6.5 and above. Make sure the selected stack is installed in a way that Run:ai can query, and verify that service discovery, authentication, and network policies permit the required metrics traffic. If your organization runs a forked or older monitoring deployment, compare its actual versions and APIs with the matrix instead of assuming feature-level equivalence.
Rank #2
- Part number 900-53651-2500-000 and model: P3651
- This is the 2 slot version for when there is no empty slots between 2 slot cards. If you have one or more empty slots between the cards or the cards are 3 slot this NVLink will not work. See the attached images showing the card layout.
- NVLink 3.0 for any brand of RTX Ampere model graphics cards: 3090, A30, A40, A100 / H100 (Requires three NVLinks), A800, A4500, A5000, A5500, A6000
- This is the same as PNY part number: NVLAMP-2SLOT-BSP and RTXA6000NVLINK-KIT
- This is the same as Dell part number: 0RWJ7Y
GPU fractions
GPU fractions have a narrower gate than ordinary whole-GPU scheduling. They require GPU Operator 26.7.1 or later, a CUDA driver at r615 or later, and a disabled NRI integration. Set the GPU Operator configuration value cdi.nriPluginEnabled=false before enabling fractional allocation. If any one of these conditions is missing, plan for whole-GPU allocation or correct the node configuration first.
RDMA and Multi-Node NVLink
Nodes that use RDMA or Multi-Node NVLink need NVIDIA Network Operator. Applicable Multi-Node NVLink configurations also use Dynamic Resource Allocation (DRA). The prerequisite baseline for Multi-Node NVLink is Kubernetes 1.32 or later and GPU Operator 25.3 or later; your Run:ai v3.0 cluster still has to satisfy its supported Kubernetes and operator ranges as well.
These features add a hardware and topology dependency: every participating node must expose the expected interconnects, drivers, and network resources. Do not enable the scheduling feature merely because the GPU model is present; confirm that the node image, firmware, operator versions, and resource APIs are aligned.
Node, storage, and access prerequisites
Operating system, runtime, and time
- Use a Linux distribution supported by both your Kubernetes or OpenShift release and NVIDIA GPU Operator.
- Configure the supported container runtime: containerd for upstream Kubernetes, or CRI-O for OpenShift.
- Synchronize every node with NTP. Consistent time prevents certificate validation, token, lease, and metrics problems that can look like unrelated control-plane failures.
- Verify x86 or ARM support across the node image, GPU Operator components, Run:ai images, and your own workload containers before mixing architectures.
Permissions and storage
Installation requires cluster-admin permissions. A restricted service account may be appropriate for day-to-day operation, but it cannot replace the elevated permissions needed to create or update cluster-wide resources during installation.
Configure a default storage class for the control plane. Run:ai worker nodes also need shared storage reachable from every worker that may run a scheduled workload. Validate access from each node, capacity, performance, identity mapping, and failure behavior; a volume that is mounted on one worker but unavailable on another is not shared storage for scheduling purposes.
Load balancing, ingress, DNS, TLS, and egress
Prepare a load balancer and ingress path before deployment. The cluster needs a DNS-resolvable fully qualified domain name (FQDN) for the Run:ai services and valid TLS certificates covering the names clients will use. Decide where TLS terminates, how certificates are renewed, and whether internal service-to-service traffic is subject to additional policy.
Rank #3
- Video/Sound Cards
- Passive Cooling
Firewall and registry allowlists must permit access to the required container registries and critical services. In a restricted network, identify proxy or mirror settings, outbound destinations, certificate chains, and image-pull credentials in advance. A chart can install successfully while its pods remain unhealthy if they cannot pull an image, resolve a service, reach the control plane, or contact the monitoring endpoint.
Guided installation versus manual Helm preparation
Guided installation
The guided path can install or configure several prerequisites for you. It is useful when the cluster has conventional permissions and networking and you want the installer to create supported components. Even then, confirm your Kubernetes or OpenShift version, GPU model, operator compatibility, storage, DNS, certificates, and registry access first; automation cannot make an unavailable provider release or unsupported GPU architecture compatible.
Manual Helm installation
Manual Helm installation assumes that infrastructure and prerequisites are already ready before charts are applied. Use it when your platform team controls operator lifecycles, ingress, certificates, storage, or air-gapped image distribution. Record the exact chart, operator, driver, and monitoring versions, and keep the values used for runtime, ingress, FQDN, TLS, registries, and storage under version control. Apply Run:ai only after the GPU Operator and monitoring stack report healthy status.
A pre-installation validation sequence
- Freeze the target release set. Write down the Run:ai release, Kubernetes or OpenShift version, runtime, node architecture, GPU model, GPU Operator version, driver version, Prometheus or Kube-Prometheus Stack version, and any Network Operator or DRA components.
- Check the platform matrix. Confirm that the exact Kubernetes or OpenShift minor release is listed for Run:ai v3.0. For EKS, GKE, AKS, OKE, or RKE2, verify that the provider offers that release and node image where you operate.
- Validate GPU nodes. Confirm that the GPU Operator version supports the exact data-center GPU and that drivers are healthy on every worker intended for scheduling. For fractions, verify operator 26.7.1 or later, CUDA driver r615 or later, and
cdi.nriPluginEnabled=false. - Validate monitoring. Confirm Prometheus 3.13.1 or Kube-Prometheus Stack 88.6.5 and above, service discovery, credentials, and network reachability.
- Validate special networking. If workloads require RDMA or Multi-Node NVLink, install and test Network Operator; for applicable NVLink designs, confirm DRA and the Kubernetes 1.32-plus baseline.
- Validate node foundations. Check Linux support, container runtime, NTP synchronization, cluster-admin access, default storage class, and shared storage from every worker.
- Validate external paths. Test load-balancer provisioning, ingress routing, FQDN resolution, TLS certificate coverage, firewall rules, registry authentication, and access to critical services.
- Install and observe. Deploy through the guided or manual path, then inspect controller, operator, scheduler, metrics, and worker pods before submitting production workloads.
Unsupported and conditional deployment cases
vGPU
Run:ai supports GPU passthrough only; vGPU is unsupported. A virtual machine or cluster design that exposes virtual GPU profiles therefore does not satisfy the documented hardware model, even if the guest operating system can load an NVIDIA driver.
NVIDIA DGX Spark, Jetson, and workstations
NVIDIA DGX Spark, NVIDIA Jetson, and workstation systems are outside the supported Run:ai v3.0 scope. Use supported data-center GPU hardware and verify the exact model against the GPU Operator list.
Framework compatibility
Cluster compatibility does not guarantee application compatibility. CUDA, PyTorch, TensorFlow, NVIDIA NIM, and other frameworks maintain their own GPU-architecture, driver, and library requirements. Check those matrices for the image and workload you plan to run. A cluster can pass every Run:ai prerequisite while a particular framework image fails because its CUDA build does not support the target architecture.
Rank #4
- CUDA Cores: 4608 / NVIDIA Tensor Cores: 576 / NVIDIA RT Cores: 72
- GPU Memory: 24 GB GDDR6 with ECC / Bandwidth: 624 GB/Sec
- System Interface: PCI Express 3.0 x16
- Four DisplayPort 1.4 Connectors
- 3D Stereo Support with Stereo Connector
Common failure modes and corrective actions
| Symptom | Likely gate | Correction |
|---|---|---|
| GPU workers show no schedulable devices | GPU Operator, driver, or unsupported GPU model | Compare the exact operator version and GPU model with the operator support list; inspect node-level driver and device-plugin health. |
| Fractional-GPU configuration is rejected or ineffective | Old operator or driver, or NRI still enabled | Use GPU Operator 26.7.1 or later, CUDA driver r615 or later, and set cdi.nriPluginEnabled=false. |
| Pods cannot start on another worker | Storage is local rather than shared | Provide shared storage reachable from every Run:ai worker and test failover between nodes. |
| UI or API is unreachable | Load balancer, ingress, DNS, TLS, or firewall configuration | Trace the FQDN from a client, verify certificate names and renewal, then check ingress and allowlists. |
| Images remain pending | Registry egress, credentials, or blocked critical service | Allow required registry destinations, install image-pull credentials, and verify proxy or mirror settings. |
| Metrics-based scheduling is incomplete | Prometheus version, discovery, authentication, or network policy | Use the listed Prometheus or Kube-Prometheus Stack version and verify that Run:ai can query the metrics endpoints. |
| Installation fails with permission errors | Insufficient RBAC | Run installation with cluster-admin permissions, then reduce permissions only for operational accounts where appropriate. |
| Interconnect workloads cannot initialize | Missing Network Operator, DRA, or topology support | Recheck RDMA or NVLink hardware, Network Operator, DRA applicability, Kubernetes 1.32-plus, and GPU Operator 25.3-plus requirements. |
How to choose between candidate clusters
When several platforms are available, score each one against the complete stack rather than choosing by Kubernetes brand alone:
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
- Release coverage: Is the exact Kubernetes or OpenShift version listed for the Run:ai release?
- Runtime and architecture: Does it provide containerd or CRI-O as required, with supported x86 or ARM node images?
- GPU path: Can you use physical GPU passthrough on a data-center model supported by the selected GPU Operator and driver?
- Operator lifecycle: Can you pin and upgrade GPU Operator, Prometheus, Network Operator, and DRA without the provider replacing them unexpectedly?
- Storage: Is there a default control-plane storage class and shared worker storage with acceptable performance and availability?
- Network exposure: Are load balancing, ingress, DNS, TLS, registry egress, and firewall changes available to the platform team?
- Topology features: If you need RDMA or Multi-Node NVLink, can the service expose the required hardware and resource APIs?
- Application layer: Do the CUDA and framework matrices support your intended images and GPU architecture?
The best candidate is the one that passes every applicable row with versions you can maintain, not necessarily the service with the newest control plane.
Frequently asked questions
Can I infer support from a different Run:ai major or minor release?
No. Check the matrix for the exact Run:ai release you are installing. OpenShift, managed-cloud, operator, and driver support can change between releases.
Is a supported Kubernetes version enough to make a managed cluster supported?
No. The provider must offer that version and the associated GPU node, storage, load-balancer, ingress, and networking capabilities in your environment.
Does Run:ai publish a universal minimum GPU model?
No. Verify the exact data-center GPU with the NVIDIA GPU Operator list and then check CUDA and framework support for the workload.
Recommended Free Tools
Can I use vGPU if the guest sees an NVIDIA device?
No. The documented deployment model is physical GPU passthrough; vGPU is unsupported.
The Bottom Line
For Run:ai v3.0, start with Kubernetes 1.35–1.37 or OpenShift 4.20–4.22, the required runtime, NVIDIA GPU Operator 25.10–26.7.1, Prometheus, supported physical data-center GPUs, and complete storage and network foundations. Treat GPU fractions, RDMA, and Multi-Node NVLink as separate feature gates, and reject vGPU, Jetson, DGX Spark, and workstation designs before installation.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
