Why NVIDIA Chose Kamaji for the DOCA Platform
Kamaji runs the Kubernetes control planes of BlueField DPU clusters as pods in the DOCA Platform Framework's host cluster.
REDUCE THE COST OF RUNNING KUBERNETES AT SCALE
- Company
- NVIDIA
- Website
- nvidia.com
- Industry
- Accelerated computing / Data centre infrastructure
- IT environment
- BlueField DPUs (ARM) with the DOCA Platform Framework (DPF)
- Clastix offering
- Kamaji
- Use case
- Hosted control planes for DPU clusters
AT A GLANCE
Challenge
Kubernetes control planes are heavyweight, and running one per DPU cluster multiplies cost and operational overhead on ARM-based edge hardware.
Solution
In the DOCA Platform Framework, Kamaji provisions a TenantControlPlane for every DPU cluster, running etcd and kube-apiserver as pods in the host cluster.
Outcome
Resource-efficient, resilient control planes for ARM-based BlueField DPU clusters, managed centrally from one host cluster.
SUMMARY
Success story overview
NVIDIA's BlueField Data Processing Units offload infrastructure workloads — networking, storage and security — from the host CPU, combining multicore ARM processors, high-speed network interfaces and hardware acceleration on the NIC. DOCA is the software framework for building on them, and the DOCA Platform Framework (DPF) is where Kubernetes comes in.
Kubernetes control planes are complex and resource-hungry, and in edge environments or with dozens of clusters that overhead multiplies. DPF answers it with a dual-cluster architecture: a host cluster where Kamaji resides, and DPU clusters made of the BlueField DPUs themselves. When a new DPUCluster is created, DPF's Kamaji Cluster Manager provisions a corresponding TenantControlPlane, deploying components such as etcd and kube-apiserver as pods within the host cluster.
Centralised control planes mean no heavy infrastructure per cluster, Kamaji's pod-based model scales to hundreds of control planes, and it runs natively on ARM — which is what lets data centres deploy and manage high-performance BlueField DPU clusters with resilience and operational efficiency.
KEY REQUIREMENTS
- Many control planes hosted on shared infrastructure
- ARM compatibility for BlueField DPUs
- High availability without dedicated VMs or bare-metal nodes per control plane
- Integration with DOCA's security and monitoring stack
KEY OUTCOMES
- Centralised control planes, with no heavy infrastructure per DPU cluster
- A pod-based architecture that scales to hundreds of control planes
- Production-grade resilience, with multi-tenancy and HA in mind
- Native management of ARM-based BlueField environments
Building a multi-tenant Kubernetes platform for your organisation?
We've built this architecture with organisations across the globe. We can help you build it too.
No commitment required. Start with a conversation.
MORE CUSTOMER STORIES
