Pure Storage has debuted FlashBlade//EXA™, said to be the industry’s highest performing data storage platform engineered for the most demanding requirements of AI and High-Performance Computing (HPC). While legacy approaches to data storage have held back AI potential, FlashBlade//EXA breaks the metadata bottleneck with a proven architecture based on FlashBlade, built for high concurrency and the massive amounts of metadata operations typical of large scale AI and HPC workloads. In preliminary testing, FlashBlade//EXA is projected to deliver more than 10 terabytes per second read performance in a single namespace - setting a new bar as the industry’s highest-performing storage solution.
The FlashBlade//EXA architecture scales data and metadata independently; provides near unlimited scale with off the shelf, third-party data nodes that enable highly scalable multi-dimensional performance; and reduces complexity in deployment, management, and scaling through the use of standard protocols and networking.
Driving a Paradigm Shift in Storage
More powerful GPUs have increased the pace and scale that foundational AI models can be trained. The explosive growth in model size and sophistication is driving a paradigm shift in storage requirements, where solutions must seamlessly keep up with the computational intensity and volume, and variety in data demands of AI and HPC.
Legacy storage systems were not designed to meet modern AI requirements. When applied to large-scale AI and HPC, they face critical limitations with parallel and concurrent reads and writes, metadata performance, ultra-low latency, asynchronous checkpointing, and predictable, high throughput.
A modern storage solution must provide a massively parallel, disaggregated architecture to deliver flexibility at scale – ensuring that storage helps accelerate the pace of AI.
A Modern Storage Architecture for AI and HPC Workloads
Legacy high-performance storage architectures were optimized for traditional HPC environments, with more predictable and regular workloads and a focus on raw performance scaling. Today’s AI workloads are complex and multi-modal — including text, images, videos — being processed simultaneously by tens of thousands of GPUs. This dramatic shift demands advanced metadata optimization alongside massive performance scaling to efficiently manage diverse data types and high concurrency.
FlashBlade//EXA is purpose-built for the challenges of AI workloads with unmatched performance and metadata management. Its disaggregated, massively parallel architecture enables storage flexibility at scale. Enterprises can adapt to evolving multimodal models, optimize reliability, and eliminate idle time to accelerate AI model training and inference while improving GPU utilization. Additionally, the combination of Pure Storage’s metadata engine and Purity operating system with cost-efficient, off-the-shelf data nodes enables enterprises to achieve an unparalleled price-to-performance ratio.
FlashBlade//EXA is built on a decade of Pure Storage innovation to provide transformational storage for AI and HPC environments. FlashBlade//EXA will:
Deliver industry-leading performance at scale, leveraging Pure Storage’s proven metadata capabilities to maximize AI pipeline efficiency and minimize delays during training and inference workflows.
Enabling multidimensional performance with massively parallel processing and scalable metadata IOPS to support high-speed AI requirements. Providing 10+ terabytes per second in a single namespace, delivering the industry’s best performance.
Reduce management complexity by eliminating metadata bottlenecks
Providing high metadata performance, availability and resiliency to handle massive AI datasets without manual tuning or additional configuration.
Accelerate AI innovation with a highly configurable and disaggregated architecture
Powering the evolving AI and HPC landscape with industry standard protocols;
Incorporating high-speed NVIDIA ConnectX NICs, Spectrum switches, LinkX cables, and accelerated communications libraries.
FlashBlade//EXA is expected to become available in summer 2025.
