Pure Storage has announced FlashBlade//Exa, which aims at artificial intelligence (AI) and high-performance computing (HPC) workloads that demand extremely high throughput to graphics processing units (GPUs). That will serve customers between large enterprise users of AI and the hyperscalers.
At the same time, FlashBlade//Exa has also introduced a new architecture to a Pure product line, one in which metadata and bulk storage are disaggregated with different hardware and protocols in use.
All of which is in line with Pure’s orientation towards architectures used by the hyperscalers, and comes hot on the heels of last week’s revelation that Meta is the mystery hyperscaler that decided to buy Pure’s Direct Flash Modules (DFMs) for its own systems (see below).
According to Patrick Smith, field chief technology officer at Pure Storage, Exa addresses challenges in storage for AI that include GPU utilisation, inconsistent performance generally, all specifically with metadata, scalability and management complexity.
Exa aims at a performance level somewhat higher than current FlashBlade products, targeting AI factories and GPU-as-a-service providers such as Coreweave, Tenstorrent, DataCrunch and Foundry, as well as research labs, HPC users and sovereign cloud projects. All of which, Pure said, have performance needs in the 1TBps (terabytes per second) to 50TBps throughput range, with 100PB (petabytes) to multiple exabytes of capacity and support for thousands to tens of thousands of GPUs.
FlashBlade is Pure’s fast file and object family, although Exa appears to be file access-only for now.
“It’s next level in comparison to the FlashBlade S500,” said Smith, citing FlashBlade//Exa performance figures of greater than 10TBps read performance in a single namespace, 3.4TBps throughput per rack, and an increase of 20 times in the number of files handled under single namespace.
The novel architecture – for Pure – that lays the ground for the new product, is disaggregation between the metadata and bulk storage data nodes. Metadata is stored on FlashBlade nodes – ie with controller hardware – and connects to customers’ compute cluster via NFS v4.1 parallel file access and TCP. Meanwhile, data nodes connect via Network File System (NFS) v3 (not parallelised) and Remote Direct Memory Access (RDMA).
For the first time, Pure will offer this with Pure-recommended network interface cards (NICs) in customer-specified commodity non-volatile memory express (NVMe) storage servers, but later this year, Pure DFMs will be available for use with FlashBlade//Exa.
As mentioned, this is the first time Pure has released a product without its own DFM capacity, but according to Smith, a decision was forced by “acceleration in the AI [artificial intelligence] landscape, increased demand and especially increased scale”.
“And so, coming out with a platform that allows customers to meet those scale demands in terms of performance and capacity is something we felt we shouldn’t wait on,” he added.
This disaggregation of metadata storage and bulk storage, as well as the independent supply of its flash modules, is in keeping with recent developments that saw it unveil Meta as a hyperscaler customer for Pure’s DFMs.
Around the turn of the year, Pure announced Kioxia and Micron as quad-level cell (QLC) flash chip providers for DFM modules for supply to “a hyperscaler” customer. That customer has now been revealed as Meta, which has gone public with a blog post detailing a shift from hard disk drives to QLC flash.
That is for workloads that suit QLC’s performance profile of highly sequential data and infrequent/low-intensity writes due to its low write endurance, and because QLC is “not yet price competitive enough for a broader deployment”.
General availability of FlashBlade//Exa will be in summer 2025. Also planned for later this year are S3 object storage access via RDMA, Nvidia certification and Pure Storage Fusion integration.
