Clusters of GPUs or TPUs are often used in machine learning and these require enormous bandwidth of data (e.g. up to 14TB/s for a NVIDIA DGX-2 cluster of 16 GPUs). Internal NVMe drives can only supply a fraction of this bandwidth. However, fabric attached NVMe arrays with EXTEN software and multiple parallel 100GB NICs can deliver the peak and sustained bandwidth needed at consistent latency to ensure GPU clusters are never idle.

Storage I/O is a critical bottleneck as HPC moves into the petaflop range. Upgrading local SAS flash to NVMe in the servers at scale is error prone, expensive and inefficient. Instead, customers are turning to fabric connected NVMe arrays. EXTEN software on industry standard target servers provides the IOPS, latency and capacity needed by supercomputers for streaming or burst buffer use cases at a much lower cost and higher resilience than local NVMe.

Global e-commerce companies must deliver increasingly rich content and better context within the transaction window of the user, all at scale across millions of users. Fabric-attached NVMe arrays with EXTEN software on industry-standard servers provide not only the required performance but also composability so that storage resources can be dynamically allocated to multiple simultaneous workloads. These hyperscaler customers are also able to leverage the increased efficiency of EXTEN powered storage to reduce overall costs while maintaining high performance.

Increasingly, enterprise data centers must support I/O intensive workloads such as AI/ML, high speed analytics, and microservice-based next-generation applications. These workloads, unlike legacy ones, are dynamic, unpredictable and have stringent latency requirements. EXTEN software defined storage solutions provide composable data paths that allow the dynamic allocation of bandwidth, latency, IOPS, and/or capacity to provide deterministic performance for multiple simultaneous workloads.