300TB NVMe over Fabric Storage Appliance

Maximizing GPU usage of your DGX systems throughout your AI journey NVIDIA’s DGX range has helped shape the AI landscape and changed future possibilities. However, the DGX range has limited internal space for NVMe flash storage, an essential element for performance and overall capability.

PNY 3S-2400 AI Optimised Storage Server creates a central pool of ultra-low latency NVMe which can be shared amongst one or multiple DGX servers. Providing each DGX with the ideal level of resource without the need for upfront over investment.

Simply connected via NVIDIA compatible EDR InfiniBand / 100Gbe, the unique RDMA protocol ensures the NVMe resource is seen and performs as if it were internal to the DGX.

Ensuring your project’s funds are better spentToday’s AI servers consume and analyse data at much higher rates than traditional storage solutions an deliver. Resulting in low GPU utilisation and dramatically extending training times as well as decreasing productivity. PNY 3S-2400 has been developed from the ground up for AI workloads and NVIDIA DGX optimisation. Delivering ultra-low latency and tremendous bandwidth at a price which allows more investment to be made on GPU resource and less on expensive, slower storage.

Excelero delivers low-latency distributed block storage for web-scale applications. NVMesh enables shared NVMe across any network and supports any local or distributed file system. The solution features an intelligent management layer that abstracts underlying hardware with CPU offload, creates logical volumes with redundancy, and provides centralised, intelligent management and monitoring.

Applications can enjoy the latency, throughput and IOPs of a local NVMe device with the convenience of centralised storage while avoiding proprietary hardware lock-in and reducing the overall storage TCO. NVMesh features a distributed block layer that allows unmodified applications to utilize pooled NVMe storage devices across a network at local speeds and latencies. Distributed NVMe storage resources are pooled with the ability to create arbitrary, dynamic block volumes that can be utilised by any host running the NVMesh block client.

Help your team be more productive by taking full advantage of the DGX capabilityUnlike most storage solutions where the initial investment dictates future growth and performance, often resulting in the need to overspend for potential future growth.

PNY’s NVmesh design can scale in stages to suit your project without any limitation. Just purchase the capacity and performance you need today and feel secure that as you scale, so can your capacity and performance.

The special 3S-2400P series models provide enhanced fault tolerance compared to the standard 3S-2400 models.

The configuration is based upon Erasure Coding, a scalable RAID-Like protection which provides various levels of resilience. The recommended configuration, including any additional scale-out units, is x + 2, meaning x data drives and 2 parity drives. In this configuration the data and parity are distributed in blocks and not just on one drive.

For example, with 12x 15TB drives a single S3-2400P unit has a raw capacity of 180TB.

However, with Erasure Coding Protection enabled and the equivalent of 2 drives’ capacity being used for Parity the useable capacity drops to 150TB. This is just one example of the many configurations supported by the 3S-2400P. Our technical team can advise on the optimum configuration for your requirements and budget.

*6500IMGsec is the maximum efficiency for a DGX1. Additional DGX’s will increase the results
**RAID5 and RAID6 based upon 15TB configurations, available in single or multiple nodes