Hybrid Cloud Software Defined Data Center

Hammerspace unveils integrated storage system for AI and HPC workloads

Hammerspace is introducing an integrated storage system for artificial intelligence and high-performance computing workloads, combining its data-management software with a new high-density flash platform. The company says the system is designed to deliver parallel data access for GPU clusters while using less rack space and avoiding the expense of simply adding conventional all-flash storage capacity.

The Hammerspace Data System brings together data unification, intelligence, orchestration, acceleration and storage. Customers can buy those capabilities with Hammerspace software, metadata nodes and high-performance storage as a system installed and supported by the company. Hammerspace also offers its software for deployment on third-party hardware and leading cloud platforms, allowing organizations to work with existing infrastructure.

Open Flash Platform hardware

At the center of the new hardware offering is Hammerspace OFP, or Open Flash Platform. The company describes OFP as a disaggregated storage design in which directly accessed SSD shelves are controlled by data processing units, or DPUs. The system uses parallel NFS (pNFS) to support parallel data access. Hammerspace says its approach separates metadata from the data path and uses GPU-server-local NVMe storage as a shared Tier 0 acceleration layer.

The OFP enclosure occupies one rack unit and contains five storage sleds. Each sled has a 64-core Arm CPU serving as a DPU, eight SSDs and two 400 GbE ports. Across the enclosure, that amounts to 40 drives and 10 400 GbE ports. The drives are supplied in two removable packs and can use TLC flash for speed or QLC flash for capacity.

With 245.76 TB QLC SSDs, Hammerspace specifies up to 9.83 PB of raw capacity per enclosure. It also lists peak bandwidth of 250 GB/s for reads and 350 GB/s for writes. These are company-provided specifications; the announcement did not include independent benchmark results.

Hammerspace gives a KV-cache example to illustrate how the system could be deployed alongside GPU infrastructure. It says 12 OFP enclosures and a one-rack-unit metadata server could support eight GPU racks, each containing 64 AMD MI355X accelerators. The company contrasts that configuration with what it says would be a four-rack, 156-node requirement for a competing all-flash system. Those figures are Hammerspace’s comparison, rather than results from a disclosed comparative test.

Five capabilities in one system

The Data System’s software layer is intended to make information available across storage systems, sites and clouds without requiring disruptive migrations. Hammerspace says applications can access unified data through standard protocols, while its data-intelligence capabilities catalog and query information across distributed infrastructure.

Orchestration policies can place data across storage systems, data centers and clouds, the company says, reducing the need for manual staging when workloads run in different locations. For acceleration, GPU-server-local NVMe can serve as Tier 0 shared data storage, while the storage layer combines OFP flash with existing or new third-party flash and hard-disk systems. The stated goal is to scale performance and capacity without creating separate data silos.

For AI workloads, one target is KV cache data displaced from a GPU’s high-bandwidth memory (HBM). Hammerspace says its Tier 0 design can make such data available across GPUs in a cluster. This differs from Nvidia’s described Tier 3.5 approach, which uses external NVMe storage and BlueField-4 DPUs to move data toward GPU HBM. Hammerspace software does not support BlueField-4, according to the supplied product details. The relative performance of the two approaches remains unproven without benchmark testing.

Availability and deployment options

Hammerspace said OFP is available for quotation and is expected to ship in the first half of 2027. The full Data System is intended for customers seeking a packaged system, while the software-only option is aimed at organizations extending deployments across their own hardware or cloud environments.

Founder and CEO David Flynn framed the product around improving performance efficiency, including throughput relative to power use, footprint and cost. The system’s real-world advantage will depend on workload behavior and measured performance, particularly for shared KV-cache access. Hammerspace’s published capacity and bandwidth figures, along with its GPU-rack example, provide an initial outline of the design but do not establish how it will compare with alternative architectures in production.