FlashBlade//S with NVIDIA DGX SuperPOD Solution Integration

FlashBlade connects to a NVIDIA DGX SuperPOD environment via sixteen 400GbE ports. This allows the FlashBlade to seamlessly scale from a minimal configuration into multiple chassis, blades, and DirectFlash Module configurations in a simple, non-disruptive manner.

NVIDIA DGX SuperPOD clusters are transforming how data is processed and analyzed, especially when integrated with advanced storage solutions. These storage systems play a crucial role in managing the large amounts of data that AI applications require, enabling seamless access and sharing for distributed computing environments. This combination allows for efficient collaboration among nodes within a compute cluster, ensuring that AI models can be trained on large datasets without delays.

The integration of Everpure FlashBlade with NVIDIA DGX SuperPOD compute clusters significantly enhances these capabilities. For example, in healthcare, AI can analyze extensive medical records and imaging data, with centralized storage providing quick retrieval and processing. In the financial sector, these systems support the storage of historical market data, allowing AI models to rapidly analyze trends and make predictions. Additionally, industries like media and entertainment benefit from this integration by managing large volumes of video and audio data for AI-driven content creation and editing.

Ethernet networking is a cornerstone of enterprise data centers due to its low total cost of ownership (TCO), extensive interoperability, and proven reliability. When FlashBlade//S is connected via Ethernet to the DGX storage fabric, it delivers exceptional performance for large AI clusters. Additionally, the use of RDMA over Converged Ethernet (RoCE) further enhances this setup by minimizing latency and offloading CPU workloads, resulting in faster data transfers and improved overall efficiency for demanding AI applications.

As the demand for AI-driven insights continues to grow, the reliance on powerful compute clusters alongside efficient storage solutions will expand, paving the way for innovation across diverse AI fields.

Figure 1. SU NVIDIA DGX SuperPOD with Everpure FlashBlade//S

FlashBlade NVIDIA DGX SuperPOD Networking

FlashBlade integrates effortlessly with NVIDIA DGX SuperPOD clusters using established Ethernet connectivity. It further enhances performance by supporting RDMA over Converged Ethernet (RoCE) communication between DGX systems and FlashBlade, ensuring low-latency data transfer. Connecting FlashBlade to a pair of SN5600 switches follows a standard installation process that has been successfully implemented in thousands of installations worldwide.

A FlashBlade array is composed of four key components: XFM modules, chassis, blades, and Direct Fabric Modules (DFMs). The XFM modules aggregate all traffic from the SN5600 switches to the blades, as illustrated in Figure 2.

Figure 2. NVIDIA DGX SuperPOD Storage Fabric Network Integration

The chassis houses the blades and associated DFMs, facilitating connectivity among the blades and to the broader network. Each blade can accommodate up to four DFMs and serves as an endpoint for client communications. Importantly, all storage provided by the DFMs is accessible to every blade within the cluster, with no file system constraints on individual blades. This architecture allows any blade in the array to access and service client requests as needed, maximizing efficiency and resource utilization.

The FlashBlade NVIDIA DGX SuperPOD Components and FlashBlade DGX SuperPOD Accessory Kit BOM can be found in Appendix A.

FlashBlade Network Connectivity

Aggregate network connectivity to the NVIDIA DGX SuperPOD storage fabric is provided by dual FlashBlade XFM8400 switches, shown in Figure 3 as ports 21-28. Each XFM uplinks to the SN5600 switches with eight connections each for a total of 16x400Gb/E connections.

Figure 3. XFM port map

Aggregate data traffic connections to the array are passed to the chassis and then blades within each chassis, fully distributing all connections across all blades in the array. All blades within the array have access to all storage available across the array and can respond to any client request.

Figure 4. Aggregate connection uplinks from XFM's to the NVIDIA DGX SuperPOD SN5600 Storage Fabric example

NVIDIA DGX SuperPOD Scaling

DGX SuperPOD to FlashBlade scaling is based on a single SU to start. As the cluster grows, the FlashBlade can be expanded to provide additional storage and IO capabilities. The table below captures sample FlashBlade deployments for a 1SU configuration; other configurations are possible. Please contact your Pure representative for complete NVIDIA DGX SuperPOD sizing options.

Figure 5. Sample Configurations for FlashBlade in a DGX SuperPOD

SU Size	FlashBlade	Capacity (raw)	Capacity (raw)	RU	Racks	Power(kW)
1	3 Chassis	2.8 PB	16	17	1	8.458
2	3 Chassis	4.4 PB	16	17	1	8.458
4	7 Chassis	10.3 PB	16	37	2	18.898
8	10 Chassis	30 PB	16	52	2	26.728

Storage Performance Guidance from NVIDIA

NVIDIA provides guidelines for DGX SuperPOD workload IO based on the following table. The numbers here are meant to represent the best case scenario and can be used to help understand the workload that a single SU would generate in terms of IO load. See the table below for more detailed information.

Figure 6. NVIDIA DGX SuperPOD Scaling Guide

Performance Workloads	Good (GBps) Text	Better (GBps) Text/Image	Best (GBps) 4k Video
Single-node read	4	8	40
Single-note write	2	4	20
Single SU aggregate	15	40	125
Single SU aggregate system write	7	20	62
4 SU aggregate system read	60	160	500
4 SU aggregate system write	30	80	250

_{Everpure, the Everpure P Logo, DirectFlash, FlashBlade, FlashBlade//S and the marks on the Everpure Trademark List are trademarks or registered trademarks of Everpure, Inc. in the U.S. and other countries. The Everpure Trademark List can be found at

purestorage.com/trademarks
.}

_{NVIDIA, DGX, and SuperPOD are trademarks and/or registered trademarks of NVIDIA Corporation in the U.S. and other countries. Other names may be trademarks of their respective owners.}