Running AI Workloads on Rack-Scale Supercomputers: From Hardware to Topology-Aware Scheduling

2026-04-07 · Source: NVIDIA Technical Blog · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cloud Computing & IT Infrastructure · Depth: Intermediate, medium

Summary

NVIDIA's GB200 NVL72 and GB300 NVL72 systems, built on the Blackwell architecture, are rack-scale supercomputers designed with 18 compute trays and high-bandwidth networking. Operationalizing these systems for AI and HPC workloads requires bridging the gap between their hierarchical hardware topology and flat scheduler abstractions. NVIDIA Mission Control addresses this by providing rack-scale control planes for Grace Blackwell NVL72 systems, integrating with workload managers like Slurm and NVIDIA Run:ai. It leverages system-level identifiers like Cluster UUID and Clique ID to reflect NVLink and IMEX domains, enabling schedulers to make topology-aware placement decisions. This ensures optimal performance, isolation, and manageability for multi-node NVLink workloads, extending support to Kubernetes via ComputeDomains and the NVIDIA k8s-dra-driver-gpu, and automating topology discovery with the open-source Topograph tool.

Key takeaway

For AI Architects and HPC platform operators deploying NVIDIA Grace Blackwell NVL72 systems, understanding and implementing NVIDIA Mission Control with Slurm, Kubernetes, or Run:ai is crucial. This approach ensures your high-bandwidth NVLink and IMEX domains are correctly exposed to schedulers, preventing performance bottlenecks from misaligned job placements and enabling efficient, scalable AI factory operations. Prioritize integrating topology-aware scheduling to maximize hardware utilization and workload performance.

Key insights

NVIDIA Mission Control and related tools bridge hardware topology with scheduler abstractions for optimal AI/HPC workload placement.

Principles

• Topology-aware scheduling is critical for multi-node GPU performance.
• System-level identifiers can encode hardware topology for software.
• Automated topology discovery scales better than manual configuration.

Method

NVIDIA Mission Control integrates with Slurm and Run:ai, using Cluster UUIDs and Clique IDs to map NVLink/IMEX domains to scheduler-consumable blocks or ComputeDomains, ensuring topology-aware job placement and resource isolation.

In practice

• Use Slurm's topology/block plugin for NVLink partition awareness.
• Implement Kubernetes ComputeDomains for multi-node NVLink workloads.
• Deploy Topograph for automated cluster topology discovery.

Topics

• NVIDIA Blackwell Architecture
• Rack-Scale Supercomputers
• NVLink Interconnect
• IMEX Shared Memory
• Topology-Aware Scheduling

Code references

• NVIDIA/k8s-dra-driver-gpu
• NVIDIA/topograph

Best for: AI Architect, MLOps Engineer, AI Engineer

Related on AIssential

• Unlock Exascale Performance on NVIDIA GB200 NVL72 with Slurm Topology-Aware Job Scheduling — Topology-aware scheduling with Slurm's topology/block plugin is crucial for maximizing NVIDIA GB200 NVL72 performance and utilization.
• NVIDIA GTC Taipei at COMPUTEX: Live Updates on What’s Next in AI — NVIDIA's latest platforms advance AI, robotics, and autonomous vehicles with significant performance and efficiency gains.
• H100 to B200: What Actually Changed — NVIDIA's Blackwell architecture, featuring the B200 GPU and GB200 NVL72 system, represents a fundamental shift from the H100, moving beyond individual GPU performance to datacenter-scale AI infrastructure.
• Lambda at GTC 2026: an early preview — Rack-scale AI infrastructure requires integrated power, cooling, and networking for optimal performance and scalability.
• Steven Sinofsky on AI PCs, NVIDIA, and the Future of Computing — The shift to AI-native hardware and local inference is fundamentally changing PC design, moving compute from cloud to device for cost-effective AI.
• NVIDIA's AI Engineers: Agent Inference at Planetary Scale and "Speed of Light" — Nader Khalil (Brev), Kyle Kranen (Dynamo) — NVIDIA GTC will feature sessions with Kyle Kranen, an architect of NVIDIA Dynamo, and Nader Khalil, from NVIDIA Brev, discussing the company's rapid growth and strategic investments in AI.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by NVIDIA Technical Blog.