In AI, No Data Is Ever Cold

Revolutionizing Storage for AI Workflows: The Open Flash Platform

The storage landscape is on the cusp of a significant transformation, driven by the escalating demands of artificial intelligence (AI) workflows. At the 2025 SNIA SDC in Santa Clara, CA, David Flynn, CEO of Hammerspace, unveiled a groundbreaking concept: the Open Flash Platform (OFP). With his rich background, including his tenure as CEO of FusionIO, the company that pioneered the SSD interface that eventually became NVMe, Flynn brought a wealth of expertise to the presentation.

The Open Flash Platform: A New Paradigm

The OFP is built upon NFS in Linux and incorporates a metadata server with data orchestration, operating in a separate control plane from the data. This innovative architecture allows for the direct movement of data from storage devices to processors, bypassing system controllers and dramatically enhancing system performance. By decoupling the data path from metadata and data orchestration, the OFP eliminates the need for traditional controllers, enabling linear scaling and higher performance to meet the stringent requirements of AI workflows.

Comparing the OFP storage architecture to traditional NAS architecture reveals significant advantages. The OFP’s design enables the elimination of bottlenecks associated with conventional storage systems, facilitating a substantial increase in performance and capacity. According to Flynn, the OFP can reduce costs by 60%, increase capacity by 10X, and operate using only 10% of the power required by conventional SSD storage.

Design and Implementation

The OFP sled, a fundamental component of the platform, consists of banks of flash memory. While conventional SSDs can be utilized, the potential for more dense packages, such as the E2 512TB Embedded NVMe SSD, offers exciting possibilities. These sleds can be stacked and equipped with a DPU and fiber interface to form a storage NIC, representing a significant leap forward in storage density and efficiency.

The implications of this technology are profound. Rows of OFP sleds can be packed into 1U shelves, which can then be stacked in racks, yielding a storage density 40X greater than conventional all-flash parallel file systems. Moreover, the OFP achieves this with 98% lower power consumption. A fully populated OFP rack can provide an Exabyte of storage, with 252 sleds delivering 200 Tbps data rates, while idle power consumption remains remarkably low at 25KW.

Collaboration and Future Directions

Hammerspace is actively collaborating with several prominent organizations, including Los Alamos National Laboratory, Linux, SK hynix ScaleFlus, and Xsight, to bring the OFP to fruition. The involvement of other leading storage companies underscores the industry’s recognition of the OFP’s potential. As AI continues to drive demand for high-performance storage, the Open Flash Platform is poised to play a pivotal role in shaping the future of data storage.

With its innovative architecture, significant performance enhancements, and environmentally friendly power consumption, the Open Flash Platform represents a critical step forward in the evolution of storage technology. As the world becomes increasingly reliant on AI and data-intensive applications, the importance of efficient, scalable, and high-performance storage solutions will only continue to grow, making the OFP a timely and groundbreaking development.