The potential for a wide variety of organizations to build and deploy AI applications to support and run various aspects of their business is significant. Accordingly, storage solution providers, including Pure Storage, continue to focus on creating products and solutions aimed at enabling such environments. Pure’s portfolio includes a range of AI-enabling capabilities; the company’s FlashBlade//S product line is certified with NVIDIA DXG SuperPOD, for example.

Right now, however, the most consequential build-out of AI infrastructure is taking place among a relatively small number of hyperscalers, GPU-as-a-service specialists, and other “AI factories” such as advanced research facilities and government agencies. This is the space Pure is targeting with its latest offering, FlashBlade//EXA.

It’s a significant announcement for a couple of reasons. First, it’s a statement of intent that Pure believes its underlying technology is capable of serving not just the needs of mainstream enterprises, but also the taxing demands of organizations building AI at extreme scale.

Second, the FlashBlade//EXA architecture emphasizes the power and extensibility of its software. Though Pure is often associated with hardware-level innovations such as its DirectFlash Modules, it is a software company at heart. FlashBlade//EXA’s architecture emphasizes this, in part because it’s a modified variant of its Purity//FB software that runs on a combination of third-party servers and SSDs, as well as Pure’s own hardware.

Pure’s logic here is based on understanding the specific needs of these types of buyers and knowing where it can add the most value. Organizations building large- and hyper-scale AI factories have typically made massive investments in compute/GPUs and related hardware; what they really struggle with is building an infrastructure that can scale to extreme levels within a cost model that makes sense.

Data and Storage Emerge as Key Bottlenecks to Achieving AI at Scale

Many of the challenges with building this type of cost model fall squarely on the storage infrastructure, which is rapidly emerging as a key bottleneck in the entire AI process. The specific issues vary depending on the stage of the AI lifecycle and the type of AI workload, but some of the most consequential problems include:

Checkpointing. Having GPUs sit idle as the system waits for checkpoints to be written to storage elongates model training times and can hugely impact the economics of the overall process.

Metadata performance. Massive structured data “maps” of underlying unstructured data need to be constructed for an AI model to function at a scale that is simply unprecedented.

The general unpredictability of resource requirements. The scale of AI workloads scale render tightly coupled approaches highly inefficient, especially when scale reaches extreme levels.

Pure’s argument is that existing HPC storage approaches, whether historically built around traditional parallel file systems or newer disaggregated approaches, are a sub-optimal fit because they were not fundamentally designed to address these challenges. Though they can meet high-scale performance requirements, they do so on the basis that the workload is well understood and predictable; even then, they are still not linearly scalable, Pure believes.

FlashBlade//EXA Overview

By contrast, Pure is designing FlashBlade//EXA with the assumption of extreme scale and unpredictability. It’s based on a disaggregated approach that separates the metadata functions from the physical data storage layer, such that both are independently scalable to extreme levels. The “Metadata Core” runs on a Pure FlashBlade array to store and manage metadata and is based on a massively scalable transactional database and key value store. This communicates with the GPU cluster, servicing metadata queries and directing the compute nodes to where the relevant data is stored.

Then, separately, the actual data is stored on a cluster of third-party data nodes, which runs a “thin” Linux-based OS with volume management and target services and is customized to work with the Metadata Core. These nodes are accessed directly by the GPU cluster at high speed via Remote Direct Memory Access (RDMA), using standard industry networking protocols (initially supporting NFS for file-level access at GA, though Pure also plans to add S3 support for object access in the future). In this way, customers can leverage their likely large pre-existing investments of 1U and 2U servers running NVMe SSDs, which can then independently scale separately from the Metadata Core as requirements change.

So what does “extreme scale” actually mean? Pure says that, while typical enterprise scale-out network-attached storage (NAS) taps out at around 1TB/sec of throughput and up to 100PB of data, FlashBlade//EXA has been designed to scale to 50TB/sec or more, supporting multi-exabytes of data and tens or even hundreds of thousands of GPUs. By any measure, these are simply enormous levels of scale.