Pure Storage has introduced FlashBlade//Exa, which goals at synthetic intelligence (AI) and high-performance computing (HPC) workloads that demand extraordinarily excessive throughput to graphics processing items (GPUs). That can serve clients between massive enterprise customers of AI and the hyperscalers.

On the similar time, FlashBlade//Exa has additionally launched a brand new structure to a Pure product line, one during which metadata and bulk storage are disaggregated with totally different {hardware} and protocols in use.

All of which is according to Pure’s orientation in direction of architectures utilized by the hyperscalers, and comes sizzling on the heels of final week’s revelation that Meta is the thriller hyperscaler that determined to purchase Pure’s Direct Flash Modules (DFMs) for its personal techniques (see beneath).

In keeping with Patrick Smith, area chief know-how officer at Pure Storage, Exa addresses challenges in storage for AI that embrace GPU utilisation, inconsistent efficiency typically, all particularly with metadata, scalability and administration complexity.

Exa goals at a efficiency stage considerably greater than present FlashBlade merchandise, concentrating on AI factories and GPU-as-a-service suppliers equivalent to Coreweave, Tenstorrent, DataCrunch and Foundry, in addition to analysis labs, HPC customers and sovereign cloud tasks. All of which, Pure mentioned, have efficiency wants within the 1TBps (terabytes per second) to 50TBps throughput vary, with 100PB (petabytes) to a number of exabytes of capability and assist for 1000’s to tens of 1000’s of GPUs.

FlashBlade is Pure’s quick file and object household, though Exa seems to be file access-only for now.

“It’s subsequent stage compared to the FlashBlade S500,” mentioned Smith, citing FlashBlade//Exa efficiency figures of larger than 10TBps learn efficiency in a single namespace, 3.4TBps throughput per rack, and a rise of 20 occasions within the variety of information dealt with beneath single namespace.

The novel structure – for Pure – that lays the bottom for the brand new product, is disaggregation between the metadata and bulk storage information nodes. Metadata is saved on FlashBlade nodes – ie with controller {hardware} – and connects to clients’ compute cluster by way of NFS v4.1 parallel file entry and TCP. In the meantime, information nodes join by way of Network File System (NFS) v3 (not parallelised) and Remote Direct Memory Access (RDMA).

For the primary time, Pure will provide this with Pure-recommended network interface cards (NICs) in customer-specified commodity non-volatile memory express (NVMe) storage servers, however later this 12 months, Pure DFMs shall be obtainable to be used with FlashBlade//Exa.

As talked about, that is the primary time Pure has launched a product with out its personal DFM capability, however in accordance with Smith, a call was pressured by “acceleration within the AI [artificial intelligence] panorama, elevated demand and particularly elevated scale”.

“And so, popping out with a platform that enables clients to satisfy these scale calls for when it comes to efficiency and capability is one thing we felt we shouldn’t wait on,” he added.

This disaggregation of metadata storage and bulk storage, in addition to the unbiased provide of its flash modules, is consistent with latest developments that noticed it unveil Meta as a hyperscaler buyer for Pure’s DFMs.

Across the flip of the 12 months, Pure introduced Kioxia and Micron as quad-level cell (QLC) flash chip suppliers for DFM modules for provide to “a hyperscaler” buyer. That buyer has now been revealed as Meta, which has gone public with a blog post detailing a shift from laborious disk drives to QLC flash.

That’s for workloads that go well with QLC’s efficiency profile of extremely sequential information and rare/low-intensity writes resulting from its low write endurance, and since QLC is “not but value aggressive sufficient for a broader deployment”.

Common availability of FlashBlade//Exa shall be in summer time 2025. Additionally deliberate for later this 12 months are S3 object storage entry by way of RDMA, Nvidia certification and Pure Storage Fusion integration.