Supported Features

This page compares feature support across different workload types in NVIDIA Run:ai. Use it to understand which scheduling, resource management, and platform capabilities are available for each workload type before selecting a workload model or submission method.

• Native workloads - Fully integrated into the platform - Workspace, Training and Inference.

• Supported workload types - A broad range of workloads from the ML and Kubernetes ecosystem enabled through the Resource Interface (RI).

• Externally submitted Kubernetes workloads - Workloads submitted outside of NVIDIA Run:ai. These workloads receive only minimal scheduling and platform capabilities.

Feature availability may vary across NVIDIA Run:ai versions and cluster deployments. Refer to this page and the linked documentation for the most up-to-date support details.

Workload Submission

	Workspace	Standard Training	Distributed Training	Inference	Distributed Inference	Supported workload types
UI
UI (via YAML)
API (Workloads v1)
API (Workloads v2)
CLI

Scheduling and Resource Management

Functionality	Workspace	Standard Training	Distributed Training	Inference	Distributed Inference	Supported workload types
Fairness
Priority and preemption
Over quota
Node pools
Bin packing / Spread
Multi-GPU fractions
Multi-GPU dynamic fractions
Node level scheduler
Multi-GPU memory swap
Elastic scaling
Gang scheduling
Network topology-aware scheduling
GB200 NVL72 and Multi-Node NVLink domains (MNNVL)
Scheduling rules

Operational and Platform Features

Functionality	Workspace	Standard Training	Distributed Training	Inference	Distributed Inference	Supported workload types
Monitoring
Workload awareness
RBAC
Workload actions (stop/run)
Rolling updates
Workload templates
Workload assets
Workload Policies

Workload awareness

Specific workload-aware visibility, so that different pods are identified and treated as a single workload (for example GPU utilization, workload view, dashboards).

Externally Submitted Kubernetes Workloads

Kubernetes workloads can be submitted outside of NVIDIA Run:ai, for example by using kubectl directly or Helm charts as part of an AI application. These workloads are scheduled by NVIDIA Run:ai and receive full monitoring support, along with a subset of scheduling capabilities.

• Supported scheduling capabilities include:

  • Fairness

  • Priority and preemption

  • Over quota

  • Node pools

  • Bin packing / Spread

  • Multi-GPU fractions

  • Multi-GPU dynamic fractions

  • Node level scheduler

  • Multi-GPU memory swap

  • Gang scheduling

• All monitoring capabilities are supported including event history, metrics and logs.