NVIDIA's new GPU goes all out with Kubernetes

What has changed

The GB200 NVL72 announced by NVIDIA isn’t just a high-performance GPU—it fundamentally changes how GPU networks are built across multiple machines. Previously, this required complex manual configuration, but now Kubernetes (a container orchestration system) can handle most of it automatically.

What is ComputeDomains?

Simply put, it’s a system that connects GPUs spread across multiple servers “safely” and “at high speed.” Integrated into NVIDIA’s DRA GPU driver, it automatically creates and manages memory access domains each time a workload (computation task) is scheduled. Security isolation and fault tolerance are also enhanced.

Benefits of Implementation

• Scalability: The entire rack becomes a unified GPU fabric, breaking through the limits of the single-node era
• Dynamic Management: Each workload gets its own independent domain, dramatically boosting resource efficiency
• Multi-Tenant Support: Multiple users’ workloads can run simultaneously without interference

Background: The evolution of GPU computing

Older NVIDIA DGX systems were limited to scaling within a single machine. The advent of multi-node NVLink (MNNVL) brought ultra-fast GPU communication across different servers. ComputeDomains implements this natively in Kubernetes, laying the groundwork for large-scale language model training and distributed inference.

What’s next

Further improvements are planned for DRA driver v25.8.0, including the removal of the single pod per node restriction and more flexible scheduling, which should further increase utilization rates. It’s the next phase for AI infrastructure.