Pure Storage has unveiled FlashBlade//Exa, a solution designed for artificial intelligence (AI) and high-performance computing (HPC) workloads. This platform is engineered to provide exceptional throughput to graphics processing units (GPUs), catering to large enterprise users and hyperscalers alike.
Innovative Architecture
FlashBlade//Exa introduces a new architectural approach within Pure’s product lineup, featuring the disaggregation of metadata and bulk storage. This allows for the utilization of distinct hardware and protocols for each type of data storage.
Targeted Performance
The FlashBlade//Exa aims to surpass the performance capabilities of existing FlashBlade products. It specifically targets performance requirements that range from 1TBps to 50TBps in throughput and supports capacities from 100PB to multiple exabytes, accommodating thousands of GPUs. Key users include AI factories and GPU-as-a-service providers, alongside research laboratories and sovereign cloud initiatives.
Enhanced Features
Patrick Smith, Pure Storage’s field chief technology officer, noted that FlashBlade//Exa tackles various challenges in AI storage, including GPU utilization, metadata management, and scalability. The performance metrics reveal that it achieves over 10TBps read performance per namespace and a throughput of 3.4TBps per rack, with the capability to manage twenty times more files under a single namespace compared to previous models.
Metadata and Bulk Storage Separation
The unique architecture separates metadata storage from bulk data nodes. Metadata is handled by FlashBlade nodes equipped with controller hardware and connects to compute clusters via NFS v4.1 for parallel file access. In contrast, bulk data nodes utilize NFS v3 and Remote Direct Memory Access (RDMA).
New Supply Chain Strategy
Pure Storage will be employing customer-specified commodity network interface cards (NICs) alongside NVMe storage servers for the first time. This transition comes as Pure’s Direct Flash Modules (DFMs) will soon be compatible with FlashBlade//Exa, allowing for an adaptable supply chain that meets customer needs more effectively.
Response to Market Dynamics
Smith indicated that the rapid evolution of the AI sector has prompted Pure to release this product without its own DFM capacity, emphasizing the necessity to meet growing demands for performance and scalability in storage solutions.
Partnerships and Future Developments
Pure Storage recently confirmed Meta as a hyperscaler customer utilizing its DFMs, following an earlier announcement of partnerships with Kioxia and Micron for quad-level cell (QLC) flash chip supply. These developments come as many organizations shift from traditional hard disk drives to QLC-based solutions, which are better suited for specific workloads but not yet cost-competitive for all applications.
Availability
The general release of FlashBlade//Exa is scheduled for summer 2025, with additional features such as S3 object storage access through RDMA, Nvidia certification, and integration with Pure Storage Fusion planned for this year.