WWT and DDN launch storage design for NVIDIA Vera Rubin AI campuses

Solving the power trade-off

World Wide Technology (WWT) and DDN have introduced a new storage reference design specifically for massive AI factory campuses. These facilities are intended to support 50 megawatts or more of power, utilizing the NVIDIA Vera Rubin NVL72 architecture. The primary goal is to address a major constraint in large-scale AI infrastructure: power consumption.

In a data center, every watt used by storage systems is a watt that cannot be allocated to GPUs. However, storage must still deliver data at high enough speeds to keep the GPUs working efficiently. This design helps operators map out their infrastructure needs by balancing storage requirements against strict site power limits.

By providing a pre-tested blueprint, the companies aim to help builders avoid late-stage design changes. As DataCenterNews Asia Pacific reported, the design allows for storage to be sized specifically for the 72-GPU rack format. This ensures that the hardware can meet the target of more than 2.5 GB/s of read throughput per GPU.

Managing the KV cache

A central technical challenge in this design involves the KV cache, which is the working memory a model builds while processing a prompt. In long-context or reasoning workloads, this cache can grow larger than the available memory on a GPU. If the system loses this cache, it must recompute the information, which adds to the GPU workload and slows down responses.

To solve this, the design uses NVIDIA CMX to create a context-memory tier between GPU memory and shared storage. This setup uses BlueField-4 STX storage processors to allow the cache to be shared across different inference nodes. By keeping the cache accessible, the system can reuse completed work instead of forcing the GPUs to recalculate it.

DDN noted that loading the KV cache from this infrastructure can be significantly faster than recomputing it. In one test involving 112K tokens, the company recorded a loading time of 2.1 seconds. This was compared to a recomputation time of 57 seconds, demonstrating a substantial performance gain for complex AI tasks.

Hardware and integration

The reference design incorporates a variety of specific components to handle different types of AI workloads. Data and checkpoints are hosted on DDN EXAScaler, while the KV cache is managed on DDN Infinia. The architecture also utilizes NVIDIA Dynamo for KV-cache placement between the different storage tiers.

Networking is handled by an NVIDIA Spectrum-X Ethernet network. This includes Spectrum-6 SN6600 switches, ConnectX-9 SuperNICs, and BlueField-4 STX storage processors. These components work together to ensure that the context-memory tier remains responsive during high-demand inference tasks.

WWT plays a critical role by integrating these various components into a single, cohesive design. Their goal is to reduce the burden on customers who would otherwise have to reconcile different vendor specifications. This allows builders to move from the design phase to actual deployment with fewer technical surprises.

Strategic planning for builders

For large-scale AI factory builders, the timing of these infrastructure decisions is critical. Many companies are currently securing land and electricity supplies while simultaneously making investment decisions about their GPU, networking, and storage needs. Because these decisions are difficult to reverse once construction begins, having a validated starting point is essential.

Felicien Mathieu, Chief Technology Officer for Middle East and Africa at WWT, noted that customers need partners who offer both performance and power-efficient designs. The reference design provides a tested path for these builders, covering everything from checkpoints to KV cache. It ensures that the facility requirements are already worked out before the first GPU is installed.

The partnership combines DDN's storage technology with WWT's testing and integration scale. Omer Asad, Chief Product Officer at DDN, stated that the collaboration provides an NVIDIA-certified storage design for Vera Rubin. This unified approach is intended to support both training and inference workloads in the next generation of AI factories.