Wednesday, April 2, 2025

Pure Targets AI Expansion Beyond Enterprises with FlashBlade//Exa

Pure Storage just unveiled FlashBlade//Exa, targeting AI and high-performance computing (HPC) workloads with a strong emphasis on high throughput for GPUs. This new offering caters to big enterprise AI users and hyperscalers.

FlashBlade//Exa comes with a fresh architecture that separates metadata from bulk storage, utilizing different hardware and protocols. This shift aligns with Pure’s strategy to serve hyperscalers and follows the recent confirmation that Meta is the mystery hyperscaler purchasing Pure’s Direct Flash Modules (DFMs).

Patrick Smith, Pure Storage’s field chief technology officer, explained that Exa tackles key AI storage challenges like GPU utilization, performance consistency, scalability, and management complexity. It aims for higher performance than existing FlashBlade models, focusing on AI factories and GPU-as-a-service providers like Coreweave, Tenstorrent, DataCrunch, and Foundry, alongside research labs and HPC projects. These users typically need performance from 1TBps to 50TBps, with capacities ranging from 100PB to multiple exabytes and support for thousands of GPUs.

FlashBlade is known for fast file and object storage, but Exa currently supports file access only. Smith highlighted that Exa delivers over 10TBps read performance in a single namespace and 3.4TBps throughput per rack, boosting the number of files managed under a single namespace by 20 times compared to the FlashBlade S500.

The innovative architecture for Exa separates metadata storage from bulk storage data nodes. Metadata resides on FlashBlade nodes with controller hardware, connecting to compute clusters through NFS v4.1 for parallel file access and TCP. Data nodes, on the other hand, connect using NFS v3 and Remote Direct Memory Access (RDMA).

This is the first time Pure offers a product without its own DFM capacity, a move driven by rapid advancements in AI and rising demand for higher performance and capacity. Smith mentioned that this platform meets those demands, pushing them to act quickly.

The separation of metadata and bulk storage and the independent availability of flash modules mirror the recent announcement of Meta as a customer for Pure’s DFMs. Earlier this year, Pure partnered with Kioxia and Micron for quad-level cell (QLC) flash chips aimed at a hyperscaler client, now confirmed to be Meta. Meta has since shared a blog post about its transition from hard disks to QLC flash, suited for workloads requiring high sequential data access with infrequent writes.

FlashBlade//Exa is set for general availability in summer 2025, with plans for S3 object storage access via RDMA, Nvidia certification, and Pure Storage Fusion integration later this year.