DynamiaDynamia
Dynamia Enterprise for HAMi

Bring enterprise GPU management
to your existing platform

Built on open-source HAMi, add GPU sharing, scheduling and runtime enhancements to Kubernetes, PaaS and AI platforms while keeping your existing workflows.

Dynamia Enterprise for HAMi · Integrate with your existing stackProduct architecture
Your Kubernetes / PaaS / AI platformKeep your existing entry points and workflows
Dynamia Enterprise for HAMi Sharing and isolation · Device scheduling · Enterprise enhancements
Kubernetes and heterogeneous acceleratorsVendor drivers · Container runtime · Device resources
ENTERPRISE READY

Let platform teams focus on their business

Enhance your existing platform’s resource management, from sharing efficiency to runtime policies.

Fine-grained sharing and isolation

Declare GPU memory, compute and device requirements. Run multiple tasks on supported devices, configure resource limits and monitor usage.

Device-dependent support

Resource pools

Organize accelerators from different vendors and models into unified resource pools, then allocate and schedule by workload demand to simplify heterogeneous resource management.

Enterprise capability

Scheduling for AI workloads

Use Binpack / Spread at node and device level, plus device type, UUID and topology to control workload placement.

Device and version dependent

GPU memory oversubscription

With UMI, idle GPU memory can be swapped back to host memory so more concurrent workloads fit beyond physical VRAM—suited to tidal colocation.

NVIDIA · Model and version dependent

Elastic GPU memory expansion

Adjust GPU memory quotas from real load; when a workload would otherwise OOM, raise the limit automatically without restarting the container.

NVIDIA · Model and version dependent

Turbo performance mode

Compute calls go straight to the CUDA Runtime—no compute hijacking. Keep only essential monitoring while VRAM limits still apply; built for low-latency workloads.

NVIDIA · Model and version dependent

Dynamic MIG

Create and remove MIG instances from predefined templates. Users declare count and memory only—MIG instance names stay transparent.

NVIDIA · Model and version dependent

Dynamic NPU

Create and reclaim vNPU instances from predefined templates. Users declare count and memory only—no need to know template names.

Subject to hardware and compatibility matrix

Workload priority

When a high-priority workload uses GPU compute, lower-priority workloads are paused and resume after it finishes—protecting critical capacity.

NVIDIA · version dependent
YOUR PLATFORM. ENHANCED.

Applications request resources
Scheduling selects devices
Runtime enforces resource limits

Kubernetes resource requests pass through admission, device selection and injection to place workloads on target GPUs. The runtime enforces supported GPU memory and compute limits.

  • Keep existing Kubernetes / PaaS entry points
  • Choose sharing, exclusive use or hardware partitioning by device capability
  • Validate performance and isolation with real workloads
Application & workload layerAI training · model inference · development workloadsRequest GPU, compute, and memory resources
HAMi EnterpriseEnterprise GPU resource management and scheduling
Scheduling selects devicesDevice selection and resource allocation
Runtime enforces limitsCompute and GPU memory control
GPU infrastructure layerPhysical GPUs · heterogeneous accelerators · vendor drivers
DEPLOY WITH CONFIDENCE

From your environment to reliable operation

Online and offline deployment, tailored to your infrastructure.

01

Confirm environment and compatibility

Verify chip model, driver, runtime, and Kubernetes version.

02

Validate representative workloads

Accept based on throughput, latency, stability, and resource usage.

03

Deploy with ongoing support

Define installation, upgrades, monitoring, and offline delivery.

BEFORE YOU START

Frequently asked questions

Do I need the full platform if I already use PaaS?

Choose Dynamia Enterprise for HAMi to integrate virtualization and scheduling into your platform. If you also need unified multi-cluster, tenant and billing management, evaluate AI Platform.

Are enterprise enhancements identical across all chips?

No. GPU memory oversubscription, elastic GPU memory expansion, Turbo and workload priority depend on the NVIDIA model and product version. Dynamic MIG and Dynamic NPU also have specific hardware requirements.

LET’S TALK

Let’s talk about your compute needs

Tell us about your devices, clusters and business goals to identify the right product, deployment approach and validation path.

Image viewer