Analyst(s): Krista Case
Publication Date: August 13, 2025
Hammerspace’s MLPerf v2.0 results show linear scaling, high GPU utilization, and 3.7x greater efficiency than its nearest competitor, positioning it as a high-performance, low-footprint AI storage solution.
What is Covered in this Article:
- Hammerspace’s MLPerf v2.0 Tier 0 benchmark results for 3D U-Net with simulated H100 GPUs.
- Demonstrated linear scaling from one to five Tier 0 nodes in throughput and GPUs supported.
- High efficiency, with 3.7x more GPUs supported per rack unit than the next competitor.
- Hammerspace’s ability to activate existing NVMe capacity for high performance and cost savings.
- Benefits of Hammerspace’s Tier 0 architecture for enterprise AI deployments in both on-premises and cloud environments.
The News: Hammerspace has released MLPerf v2.0 benchmark results showcasing the speed, scalability, and efficiency of its Tier 0 storage architecture. In the 3D U-Net test – used to simulate medical image segmentation workloads with H100 GPUs – Tier 0 delivered linear scaling from one to five nodes, hitting 420.8 GB/s throughput and supporting up to 140 GPUs with 96.4% utilization.
The results also show Hammerspace leading in efficiency, supporting 3.7x more GPUs per added rack unit of storage compared to its closest rival. This edge comes from Hammerspace’s ability to pool existing NVMe storage inside GPU servers, keeping the need for extra hardware to a minimum.
MLPerf v2.0 Results Highlight Hammerspace Tier 0’s Role in Maximizing GPU Utilization
Analyst Take: Hammerspace’s MLPerf v2.0 benchmark results show that tapping into existing NVMe storage across GPU servers can boost AI workload performance and efficiency. By delivering high throughput and GPU utilization with minimal new infrastructure, Hammerspace’s Tier 0 architecture allows enterprises to get more from their GPU investments without major capital expenses.
Linear Scaling Across Nodes
In the MLPerf v2.0 3D U-Net test, the Hammerspace architecture went from handling 28 GPUs at 85.6 GB/s on one node to 140 GPUs at 420.8 GB/s on five nodes. Average GPU utilization stayed above 94.7% across all setups, with a very low coefficient of variation (0.08%-0.14%), showing consistent, repeatable performance. This linear growth indicates that the Hammerspace Tier 0 architecture can scale without sacrificing performance – a particular value as organizations seek to expand their AI clusters.
GPU Efficiency
Hammerspace’s Tier 0 architecture optimizes the number of GPUs supported per extra rack unit of storage. In the benchmark, it handles 140 GPUs per unit – 3.7x more than the closest competitor. For data centers tight on power, cooling, or space, this means more computing power with less hardware. In the benchmark configuration, it required just one 1U metadata server (Anvil) for coordination, keeping its footprint small while maintaining strong performance. Even with a high-availability setup using two Anvils, it’s still 85% more efficient than the next-best option, making it a strong pick for organizations focused on getting the most performance per watt and rack unit.
Activating Existing NVMe for Performance Gains
Hammerspace’s Tier 0 architecture pulls together the local NVMe drives already in GPU servers into one shared, high-speed storage tier. In the test setup, each client had ten ScaleFlux CSD5000 drives, with data accessed directly over parallel Network File System (pNFS) after layout from the Anvil. This setup removes the bottlenecks of external shared storage while making use of already-installed NVMe drives. Activation can happen in hours without major upgrades, making it easy to kick off AI projects quickly.
Benefits for Enterprise AI Adoption
For businesses, the Hammerspace architecture streamlines AI storage by working with current network and hardware setups as it does not require special networking or agent software. It can deliver up to 10x the performance of traditional network storage, speeding up the creation of checkpoints, boosting GPU utilization, and cutting inference latency. Collectively, these efficiencies shorten the time it takes for AI projects to deliver results. Additionally, the ability to move data between tiers, placing it on high-speed NVMe for processing and then archiving results to lower-cost storage, adds significant flexibility in managing AI workloads for cost efficiency.
What to Watch:
- How quickly enterprises adopt Hammerspace Tier 0 in production AI workloads following these benchmark results.
- The role of Hammerspace Tier 0 in reducing total cost of ownership for AI infrastructure by minimizing additional storage purchases.
- Potential competitive responses from other HPC and AI storage vendors targeting similar efficiency gains.
- The impact of Hammerspace’s orchestration capabilities on multi-tier, multi-cloud AI deployment strategies.
See the complete press release on Hammerspace’s MLPerf v2.0 benchmark results on the Hammerspace website.
Disclosure: Futurum is a research and advisory firm that engages or has engaged in research, analysis, and advisory services with many technology companies, including those mentioned in this article. The author does not hold any equity positions with any company mentioned in this article.
Analysis and opinions expressed herein are specific to the analyst individually and data and other information that might have been provided for validation, not those of Futurum as a whole.
Other insights from Futurum:
Oracle Adds Hammerspace to Strengthen AI Storage on OCI
Can Hammerspace’s Tier 0 Unlock the Full Potential of AI GPUs?
Hammerspace Tier 0: Unlocking Greater Efficiency in GPU-Driven Computing
Author Information
Krista Case brings over 15 years of experience providing research and advisory services and creating thought leadership content. Her vantage point spans technology and vendor portfolio developments; customer buying behavior trends; and vendor ecosystems, go-to-market positioning, and business models. Her work has appeared in major publications including eWeek, TechTarget and The Register.
