Topology-Aware Workload Scheduling with NVIDIA Topograph

NVIDIA Releases Topograph to Optimize AI Workload Placement

NVIDIA has introduced Topograph, an open-source toolkit designed to automate the discovery of data center and cloud cluster topologies for AI workload schedulers. Announced September 22, 2026, the tool addresses a critical operational challenge in large-scale AI factories: inefficient GPU workload placement. By providing schedulers like Kubernetes and Slurm with an accurate, continuously updated view of the physical network fabric, Topograph aims to prevent performance degradation and reduce operational costs associated with suboptimal resource allocation, where GPUs can sit idle while waiting for data over congested network links.

Topograph functions by using a provider to discover hardware connections—either from cloud APIs or on-premises fabric management systems—and normalizing this data into a common model. An engine then translates this model into a format consumable by specific workload managers, such as Kubernetes node labels (e.g., `fabric.topograph.run/tier-0`) or Slurm configuration files. This ensures that tightly coupled distributed training jobs are placed within the closest locality domains, leveraging high-bandwidth interconnects like NVIDIA NVLink and Spectrum-X. The system is designed to dynamically regenerate its topology view when cluster changes occur, eliminating the need for manual maintenance of static configuration files.

Key Capabilities and Platform Support

• Discovery Providers: Supports major cloud platforms including Google Cloud, Lambda, and Oracle Cloud Infrastructure (OCI), as well as on-premises InfiniBand and Spectrum-X fabrics.
• Scheduler Integration: Publishes topology data for Kubernetes, Slurm, and Slinky, and works alongside the KAI Scheduler for topology-aware gang scheduling.
• Dynamic Updates: A Node Observer component watches for cluster changes and automatically triggers topology regeneration to keep scheduler data current.
• Deployment: Available as a Helm chart for Kubernetes deployments or as native packages for Slurm clusters, with simulation utilities available for testing.

For the broader AI ecosystem, Topograph represents a key piece of infrastructure automation within the NVIDIA DSX OS orchestration layer. By abstracting the physical complexity of underlying hardware, it allows organizations to operate large, heterogeneous clusters more efficiently. This standardization across multi-cloud and on-premises environments enables more fluid and performant scheduling, which is essential as AI models and their corresponding hardware clusters continue to scale in size and complexity. The tool's open-source availability on GitHub is intended to foster wider adoption and community-driven development for new provider integrations.

By abstracting physical hardware complexity into a standardized, dynamic data model, NVIDIA Topograph provides the foundational layer for treating disparate on-premises and multi-cloud GPU resources as a single, optimizable fabric for AI schedulers.