DDN Infinia Integrates with NVIDIA Inference Data Transfer System

DDN has integrated its storage platform Infinia with NVIDIA's inference data transfer system. This collaboration focuses on reducing the time GPUs spend waiting for data in large language model services. On July 23, DDN announced on its technology blog that the Infinia NIXL plugin is included in NVIDIA's NIXL 1.3 distribution and Dynamo inference container. NIXL is a library that handles data movement between GPUs, CPUs, SSDs, and object storage in distributed inference environments. The integration emphasizes the movement of KV caches that occur during the AI inference process, where KV caches are intermediate data stored by LLMs to reuse context. NVIDIA's Dynamo documentation explains that KV caches can be moved to GPU memory, CPU memory, SSDs, and network-attached storage to reduce recomputation and lower response latency. Starting from NIXL 1.3, the Infinia plugin is included in the standard package, and users only need to add the DDN client package. DDN stated that the KV cache acceleration feature can be activated through configuration without changing application code. Sven Oume, DDN's CTO, explained that Infinia directly supplies context data to the GPU. This collaboration addresses the storage and transfer bottlenecks in AI data centers, with DDN and NVIDIA improving the pathways for moving stored context data to GPUs during the inference process.