Enterprise storage hardware vendor Pure Storage has announced its brand new FlashBlade//EXA platform, building on its existing strengths and directly targeting those working on large-scale on-premises artificial intelligence where performance is a must.
Pure Storage made the announcement at this week's Nvidia global technology conference (GTC), with the new FlashBlade//EXA storage unit building on the company's existing FlashBlade hardware and the powerful simplicity of its Purity operating system.
"We are seeing many exciting use cases in AI and HPC in Australia & New Zealand and around the Asia Pacific. The FlashBlade//EXA is a perfect solution for organisations that are executing large scale AI and HPC projects and have faced metadata and performance bottlenecks with legacy storage systems. This solution empowers our channel partners to unlock new opportunities for their customers, driving innovation and growth across the region," said Pure Storage VP and CTO Asia Pacific & Japan Mark Jobbins (pictured).
While companies worldwide are making huge investments in GPU hardware the challenges of scalable, efficient, high-performance AI demand shining a light on the rest of the hardware stack to prevent bottlenecks.
This is precisely where the FlashBlade//EXA comes in; the new device has been built with high-performance computing (HPC) and large-scale LLMs and RAGs right in its sights. To manage increasingly complex workloads HPC environments moved from the traditional model of computing power and storage together, to have compute and storage independent of each other. However, this still mixed data and metadata together, and metadata has become the bottleneck holding up the AI training and inference pipelines. FlashBlade//EXA solves this by now separating data from metadata on all-flash volumes.
{loadposition david08}
Traditional HPC storage was designed for predictable, traditional workloads. Generative AI, with its vast thirst for huge, content-dense data stored in all manner of structured and unstructed forms is different. To keep tabs on such a range of disparate files and formats we need metadata - data about the data. That is, this file has a name, a filesize, is of a certain media type, and all kinds of other properties about the data without actually needing to analyse the data itself. This allows for faster sorting, selection, and location of data needed for any given task but metadata must be kept up-to-date as files change, and metadata must be read from disk just as the data is. All of this combines to mean AI pipelines stall - all the GPU grunt in the world doesn't matter when you can't feed it fast enough to keep it occupied.
This is the very challenge Pure Storage has tackled head on. The new FlashBlade//EXA decouples data and metadata nodes, architecting for scale, and all under the same, familiar Purity OS that spans the entire Pure Storage product line. Purity OS has always been good with metadata management and parallelised I/O across variable size data, and now it has stronger hardware support to take it further.
Thus, the Pure Storage brand new FlashBlade//EXA platform is rightly dubbed the all-in-one AI data platform, bringing near unlimited scale, industry leading TCO, and 10+ TB/s of multi-dimensional performance. Pure Storage says it brings 20x more files in a single namespace and a 50% reduction in installation time than the leading parallel file system. It offers over 3,.4TB/sec per rack. The metadata store is accessible using standard NFS v4.1 over TCP, and NFS v3 over RDMA for data nodes. Data nodes can hold up to 24 DirectFlash 75TB/150TB modules. The platform can be automated and orchestrated through REST APIs and brings huge performance with enterprise-grade scale and resiliency.
It's a big announcement, with customer testing ramping up through March 2025, and an expected general availability by the end of September 2025.
As per its usual pattern, Pure Storage will rapidly ramp up its roadmap following general availability, bringing S3 over RDMA, direct flash module data nodes, Fusion integration, Nvidia certification and more.
The device is impressive, but it's not for everyone; Pure sees the most likely use case as HPC environments or AI titans performing AI at scale providing a massively parallel, disaggregated architecture to deliver flexibility at scale and ensuring that storage helps accelerate the pace of AI, not hinder it.