GPUDirect Storage
Networking & InterconnectA technology that creates a direct data path between storage and GPU memory, bypassing the CPU to feed accelerators faster.
Keeping B300 GPUs supplied with data is as important as the GPUs themselves. Eleveight AI pairs its cluster with WEKA high-performance storage of 2.7 PB (up to 3.2 PB) and 3.2 PB of Dell PowerScale, with a direct storage-to-GPU path that helps avoid starving the accelerators during data-heavy training.
Overview
A GPU is only as fast as the data reaching it, and in large training runs the volume of data is immense. Conventionally, data travels from storage into CPU memory first, and only then on to the GPU, an extra hop that consumes CPU cycles and adds latency. GPUDirect Storage removes that detour, letting data flow directly from storage devices into GPU memory.
How it works
The technology establishes a direct memory path between the storage subsystem and the GPU, so that data bypasses the so-called bounce buffer in CPU memory entirely. Working alongside high-throughput parallel file systems, it lets accelerators pull from storage at high bandwidth without routing every byte through the host processor, cutting both latency and CPU overhead in the data pipeline.
Why it matters
In data-intensive training, the storage path can quietly become the bottleneck, leaving expensive GPUs idle while they wait for the next batch. Feeding accelerators directly from storage keeps them busy, which is the entire economic point of high-end hardware. As datasets grow, an efficient path from storage to GPU memory becomes as decisive as the compute and interconnect around it.
Use cases
- High-throughput training data pipelines
- Large-scale data preprocessing
- Feeding multi-GPU clusters without CPU bottlenecks
- Data-intensive scientific and analytics workloads