DDN and Nvidia are joining forces to let graphics processing units grab their own data directly, a move that could slash the latency and cost that currently choke large-scale AI workloads. The partnership targets a well-known bottleneck: GPUs often sit idle while waiting for data to travel from storage systems over congested networks. By enabling direct GPU-to-data access, the companies aim to keep those expensive chips busy longer and reduce the infrastructure bill for AI training and inference.
The data pipeline problem
Training a large language model or running real-time inference involves moving enormous datasets between storage and compute nodes. Traditional architectures route data through the CPU and system memory before it reaches the GPU, adding milliseconds of delay that compound over billions of operations. DDN, a storage specialist for high-performance computing, and Nvidia, the dominant GPU maker, are working on a solution that lets the GPU talk directly to the storage layer. The result: fewer hops, lower latency, and less wasted compute time.
What the collaboration actually does
The two companies are integrating DDN's storage software with Nvidia's GPUDirect technology. GPUDirect already allows GPUs to communicate with each other and with network adapters without CPU involvement. The new work extends that direct path to DDN's parallel file system, so the GPU can fetch and store training data without waiting for the CPU to broker the transaction. Early tests show the approach can cut data-access latency by double-digit percentages, though DDN and Nvidia have not released specific benchmarks.
As AI models grow larger and training datasets balloon, the data pipeline becomes a bigger fraction of total job time. Reducing that overhead means either faster training cycles or the ability to run more experiments within the same budget. For companies building AI infrastructure, lower latency can translate directly into lower capital expenditure—fewer GPUs needed to hit a given throughput target. The collaboration also hints at a future where storage and compute are more tightly coupled, a trend that could reshape how data centers are designed.
Neither DDN nor Nvidia has announced a timeline for when the integrated solution will be commercially available. Both companies are expected to demonstrate the technology at upcoming industry events, with early access likely for select customers running large-scale AI workloads.
