NVIDIA GPUDirect Storage and Pavilion

The combination delivers unprecedented low latency, along with all the other good stuff you’d expect.


Embarrassingly parallel applications (those that require little or no effort to separate the problem into parallel tasks) including AI and big data analytics, are leveraging the power of GPU based computing because they provide organizations with greater insights, based on more data ingested faster than ever before. 

As organizations are embracing these modern applications, the impact of IO on the success or failure of these initiatives has become increasingly clear. To get the most out of GPU computing, the IO must include more than just high throughput. Low latency is critical to keeping GPU cores active. 

Without low latency storage, AI training and inference takes longer, which means less data can be analyzed, and key insights can be missed. 

Recognizing this reality, NVIDIA has developed NVIDIA® GPUDirect® Storage (GDS). GDS provides a direct data path for direct memory access (DMA) transfers between GPU memory and storage. This enables IO to occur between GPU memory and storage without utilizing the system CPU, dramatically lowering latency.

The Pavilion HyperParallel Data Platform, which delivers the most performant, lowest latency, scalable and dense storage system partners with NVIDIA and GDS to address this key requirement. 

The Pavilion HyperParallel Data Platform supports NVMe-RoCE to deliver the performance and low latency of direct attached storage with the management and flexibility of external storage with linear scalability.

Learn more about how Pavilion works with NVIDIA and GPUDirect Storage to shatter expectations for AI/ML, Deep Learning, Big Data Analytics, and other high performance applications.

To learn more about NVIDIA GPUDirect Storage visit the NVIDIA GPUDirect site.

To discover what’s possible for your environment with Pavilion, reach out.