Hammerspace is launching its own complete storage system for AI workloads, with its own dense and accelerated all-flash storage servers, and tiered storage layers not using Nvidia’s tier 3.5 KV caching scheme.
The Hammerspace Data System combines data orchestration, data intelligence, existing storage and new high-performance flash storage in one unified system for AI and HPC. There is an option option to purchase Hammerspace’s offerings as a system to simplify procurement, deployment and overall customer experience.
David Flynn, Hammerspace Founder and CEO, said: "The race for AI performance isn't slowing down. But the next era of AI infrastructure will be defined by how efficiently we deliver that performance — maximizing throughput while minimizing power, footprint and cost.”
Hammerspace says AI demands extreme performance, but achieving it by over-provisioning SSD capacity or adding in-band storage servers drives unnecessary cost, power consumption and performance bottlenecks. it takes a different approach: separating metadata from the data path, enabling high-performance parallel I/O, and leveraging GPU server-local NVMe for Tier 0 acceleration. Combined with intelligent data orchestration across existing infrastructure, clouds and neoclouds, this architecture delivers the performance AI demands with greater efficiency per dollar, watt and rack unit.
Flynn again: “Simply adding more storage isn't the answer. AI factories simplified compute. We're bringing that same transformation to data systems by unifying intelligence, orchestration, high-performance access and storage in one system. With Hammerspace OFP, we're delivering more performance in less rack space and reducing storage infrastructure overhead, while giving organizations the freedom to put their data to work wherever AI runs.”
OFP is the Open Flash Platform and it is a group of suppliers, led by Hammerspace, which wants to replace existing all-flash storage servers with disaggregated, DPU-controlled, shelves of directly-accessed new style SSDs, using pNFS to provide low latency, high bandwidth and scalable parallel data access.
Hammerspace OFP
This is a 1 RU enclosure containing five long storage server sleds, each containing a 64-core ARM CPU (DPU), and 8 SSDs, making 40 drives in total, and 2 x 400 GbE ports. The SSDs are in two removable drive packs and can be TLC or QLC drives for speed and capacity respectively. The enclosure has a total of 10 x 400 GbE ports and up to 9.83 PB of raw capacity using 245.76 TB QLC SSDs.
As you might expect from these specs, the performance is quite high with 250 GB/s read bandwidth and 350 GB/sec write bandwidth. On a per-rack unit basis this blows away the competition and reduces the amount of rack space needed to service GPU racks. A Hammerspace example is that 12 of these enclosures and a 1 RU metadata server can support 8 GPU racks, each containing 64 AMD MI355X accelerators, on a KV cache workload. It claims a competing all-flash storage system would need 4 racks with 156 nodes in them
The savings in rack space, power and cooling will be substantial.
Hammerspace Data System
Hammerspace says its combining five capabilities into a single system:
- Data Unification — Put existing data to work for AI faster. Unifies data across storage systems, sites and clouds in minutes, without disruptive migrations. Existing applications and AI workloads access the same data through standard protocols.
- Data Intelligence — Find and activate more data. Catalogs and queries data across distributed infrastructure, connecting data intelligence directly to automated orchestration policies.
- Data Orchestration — Deliver data wherever AI runs. Automates data placement across storage systems, data centers and clouds, eliminating manual staging and enabling workloads to run wherever compute is available.
- Data Acceleration — Keep GPUs productive. Delivers high-performance parallel reads and writes, with Tier 0 using server-local NVMe as a shared data accelerator to eliminate bottlenecks and improve GPU utilization.
- Data Storage — Maximize performance and capacity efficiency. Integrates Hammerspace OFP high-performance flash with existing and new third-party flash and HDD storage, scaling performance and capacity without creating new silos.
Customers can purchase the full Hammerspace Data System as one integrated system, combining all five capabilities with Hammerspace software, metadata nodes and high-performance storage, all installed and supported by Hammerspace. They can also deploy Hammerspace software as a software-defined offering on third-party hardware or leading cloud platforms, extending their existing infrastructure without creating new storage silos.
Availability
The Hammerspace OFP storage system is available to quote now and will ship in the first half of 2027.
Comment
This Hammerspace Data System combines the SSD drives in a GPU server’s x86 CPU’s local storage into what it calls a shared tier 0 storage pool for all the GPUs in a cluster. KV cache contents evicted from a GPU’s high-bandwidth memory (HBM) will be placed here and can be reloaded into HBM if needed. The Hammerspace system will use this to store KV cache data and make it sharable across GPUs.
Nvidia’s KV caching scheme employs tier 3.5 storage, special external storage NVMe SSDs with data loaded directly into HBM. The external storage uses Bluefield-4 (BF4) DPUs which link to similar DPUs in the GPU server.
Hammerspace’s SW does not support BF4s but BF4s are Arm-based systems.
Which KV cache scheme is faster; Nvidia’s tier 3.5 or Hamerspace’s tier 0, will need benchmark testing, but both will be high speed and Hammerspace could have the edge in terms of bandwidth, bandwidth per RU, and storage density over Nvidia’s storage supplier partners.