Pure Storage introduces FlashBlade//EXA, a high-performance AI and HPC storage platform delivering 10+ TB/s read speed at scale.
Pure Storage has introduced FlashBlade//EXA, touted as the industry’s highest-performing data storage platform designed to manage the most demanding AI and high-performance computing (HPC) workloads. This new offering, built on the proven FlashBlade architecture, eliminates metadata bottlenecks that historically constrained AI’s potential. Engineered for high concurrency and adept at handling the extensive metadata demands typical of large-scale AI and HPC, FlashBlade//EXA is projected to provide over 10 terabytes per second of read performance within a single namespace—establishing a new performance standard.
FlashBlade//EXA’s foundational architecture scales data and metadata independently. Organizations can achieve nearly limitless expansion and multidimensional performance by utilizing readily available third-party data nodes. Additionally, its reliance on standard protocols and networking standards simplifies deployment, management, and scaling processes, making it a straightforward fit for various enterprise environments.
Driving a Paradigm Shift in Storage
As GPUs become increasingly powerful, accelerating the training of large AI models, data storage systems must adapt to the rising demands of computational intensity, data volume, and data diversity. Traditional storage solutions, designed initially for predictable HPC workloads, often fall short in critical areas for modern AI, such as concurrent reads and writes, metadata performance, low latency, and asynchronous checkpointing.
Rob Lee, Chief Technology Officer at Pure Storage, underscored the importance of a solution that scales data and metadata independently. He emphasized that FlashBlade//EXA’s massively parallel architecture delivers unmatched performance, scalability, and adaptability, noting that next-generation storage can now actively drive the evolution of large-scale HPC and AI.
A truly modern storage solution should offer a disaggregated, parallel design to accelerate AI at scale effectively. Organizations can effortlessly add resources to keep pace with AI’s rapid and continual transformation by shifting away from a one-size-fits-all approach.
So what is Metadata and why does it Metadata matter
Metadata refers to the information about stored data that helps in its identification, categorization, accessibility, and retrieval. In data management and storage arrays, metadata includes critical details such as file names, sizes, creation dates, permissions, and locations within the storage system. As storage systems scale and datasets grow larger, the volume of metadata also increases significantly. Without efficient metadata management, this growth can lead to performance bottlenecks, slower data retrieval, and increased latency. This is particularly problematic for AI and HPC workloads that require rapid, concurrent access to massive datasets. Effective handling of metadata is therefore crucial for maintaining high performance, reliability, and efficiency in large-scale storage environments.
A Modern Architecture for AI and HPC Workloads
Legacy high-performance storage technologies were developed to handle predictable HPC workloads focused on raw performance scaling. In contrast, today’s AI tasks require processing various data types—text, images, and video—parallel across tens of thousands of GPUs. This shift necessitates architectures that optimize metadata while providing significant performance to manage substantial concurrency efficiently.
FlashBlade//EXA addresses these requirements with a disaggregated, highly parallel design that targets performance and metadata scaling. This approach helps enterprises adapt to growing multimodal AI models, improve reliability, and reduce idle times. To further optimize AI and HPC workloads, FlashBlade//EXA integrates high-speed NVIDIA ConnectX NICs, Spectrum switches, LinkX cables, and accelerated communications libraries, ensuring seamless data movement and low-latency performance for large-scale training and inference. As a result, organizations can accelerate training and inference, maximize GPU utilization, and reduce overall time to insight. Additionally, by pairing Pure Storage’s metadata engine and Purity operating system with cost-effective, commodity data nodes, enterprises can achieve a standout price-to-performance ratio that meets the demands of large-scale AI projects.
Key Capabilities of FlashBlade//EXA
Industry-Leading Performance at Scale
FlashBlade//EXA leverages Pure Storage’s extensive experience in metadata optimization to maximize pipeline efficiency for AI training and inference. According to preliminary tests, it can deliver over 10 terabytes per second of read performance in a single namespace, setting a new industry performance threshold. This capability is driven by massively parallel processing and scalable metadata IOPS, enabling organizations to meet high-speed AI requirements seamlessly.
Reduced Management Complexity
By removing metadata bottlenecks, FlashBlade//EXA addresses one of the biggest challenges in AI workloads: efficiently handling massive datasets. High metadata performance, availability, and resiliency allow AI deployments to run without manual tuning or specialized configuration. As a result, enterprises benefit from fewer operational hurdles when managing large-scale or rapidly evolving data sets.
Accelerated AI Innovation
FlashBlade//EXA’s disaggregated architecture is highly configurable and integrates with industry-standard protocols, including NVIDIA’s accelerated communications libraries, to enable high-bandwidth AI training and inference. This deep integration ensures that enterprises maximize GPU utilization while maintaining the flexibility to scale AI workloads efficiently. This design promotes seamless evolution within the AI and HPC landscape, ensuring that organizations can quickly adapt to new data types, innovative applications, and emerging models without compromising performance or reliability.
Availability
FlashBlade//EXA is slated for release in the summer of 2025.
Engage with StorageReview
Newsletter | YouTube | Podcast iTunes/Spotify | Instagram | Twitter | TikTok | RSS Feed
