Tier 0
Hammerspace Tier 0 is a software-defined storage solution that transforms the local NVMe drives inside GPU servers into a high-performance, shared, parallel file system accessible via standard NFS/pNFS protocols.
Publisher review
Hammerspace Tier 0 is a software-defined storage solution that transforms the local NVMe drives inside GPU servers into a high-performance, shared, parallel file system accessible via standard NFS/pNFS protocols. It is designed for organizations running AI training, checkpointing, inferencing, and agentic AI workloads that require low-latency, high-throughput data access without the cost and complexity of external flash arrays. By activating the unused local storage already present in compute clusters, Tier 0 eliminates I/O bottlenecks and maximizes GPU utilization, making it suitable for enterprises, cloud providers, and HPC centers seeking to accelerate data-intensive pipelines while reducing infrastructure spend.
The solution can be deployed on-premises or in the cloud and is available across all four major US cloud platforms. Tier 0 works by aggregating server-local NVMe capacity into a unified global namespace managed by Hammerspace's Anvil metadata servers, which use pNFS layouts to direct I/O directly to the closest storage target. With the v5.2 release, Tier 0 affinitization automatically detects when a compute node hosts a local Tier 0 volume and prioritizes that volume in pNFS layouts, ensuring most I/O stays local and avoiding network round trips.
This eliminates the need for manual directory configuration or proprietary kernel patches. The system supports multi-protocol access (NFS, SMB, S3), automates data tiering to cloud object storage, and can be activated in under 30 minutes with no client or agent installation on compute nodes. In cloud deployments, Tier 0 delivers 2–2.5x higher throughput and 50% lower latency compared to native cloud file systems.
Hammerspace Tier 0 competes directly with Vast Data, Weka, and DDN with Lustre, which rely on external NVMe flash arrays and high-speed networking. Unlike these alternatives, Tier 0 repurposes existing server-local NVMe, avoiding the cost, power, and rack space of dedicated storage nodes. However, it introduces dependency on Hammerspace's Anvil metadata servers for layout coordination, and performance scales with the number of local NVMe drives per node rather than a centralized flash pool.
The solution also requires Hammerspace's full data platform for orchestration and protection, which may add licensing overhead for smaller deployments. While Tier 0 reduces the need for external storage, it does not eliminate it entirely—colder tiers on object storage are still recommended for capacity. Additionally, the affinitization feature is only available in v5.2+, so older deployments require manual configuration to achieve locality.
How it works
-
Local NVMe activation
Turns GPU server-local NVMe drives into high-performance shared storage managed by Hammerspace, eliminating the need for external flash arrays.
-
pNFS-based data access
Uses Parallel NFS standards to separate metadata from data path, enabling direct client-to-storage I/O without metadata server bottlenecks.
-
Automatic locality affinitization
Hammerspace v5.2 automatically detects local Tier 0 volumes and prioritizes them in pNFS layouts, maximizing local I/O without manual configuration.
-
No client or agent install
Compute nodes require no proprietary software or kernel patches; all functionality is based on standard NFS/pNFS protocols.
-
Multi-protocol unified namespace
Provides a single global namespace accessible via NFS, SMB, and S3, joining edge, core, and cloud regions together.
-
Automated data orchestration
Policy-based data tiering automatically moves data between Tier 0, network-attached flash, and cloud object storage while maintaining visibility.
-
Cloud performance boost
Delivers 2–2.5x higher throughput and 50% lower latency vs. cloud file systems, reducing GPU idle time and accelerating AI pipelines.
Strengths and trade-offs
Strengths
- Reduces checkpointing time for AI training by up to 10x compared to external networked storage, enabling faster model iteration.
- Delivers 2–2.5x higher throughput and 50% lower latency in cloud deployments vs. native cloud file systems, boosting GPU utilization.
- Eliminates the need for external flash storage arrays, reducing hardware costs, power consumption, and rack space by repurposing existing local NVMe.
- Activates in under 30 minutes with no client or agent installation, minimizing deployment complexity and time to value.
Trade-offs
- Performance is inherently limited by the number and speed of local NVMe drives per node, rather than a centralized flash pool that can be scaled independently.
- Requires Hammerspace Anvil metadata servers for layout coordination, introducing a single point of failure if not properly replicated.
- The affinitization feature that ensures local I/O is only available in Hammerspace v5.2+; older versions require manual directory configuration to achieve locality.
- Colder data tiers (e.g., object storage) are still needed for capacity, meaning Tier 0 does not fully replace traditional storage for large datasets.
Pricing context
Not specified in sources
Getting started with Tier 0
-
Sign up for Hammerspace
Register for a Hammerspace account through their website. Provide your organization details and verify your email to gain access to the Hammerspace management console and download links for the Anvil metadata server software.
-
Deploy Anvil metadata servers
Install and configure the Hammerspace Anvil metadata servers on dedicated nodes in your cluster. Set up at least two servers for high availability, ensuring they can communicate with all GPU compute nodes over your network.
-
Activate local NVMe storage
On each GPU server, identify the unused local NVMe drives. Use the Hammerspace management console to designate these drives as Tier 0 storage, which aggregates them into a shared parallel file system accessible via NFS/pNFS.
-
Mount the Tier 0 namespace
On your compute nodes, mount the Hammerspace global namespace using standard NFS/pNFS commands. No client or agent installation is required; simply point to the Anvil metadata server IP and the exported path to access the unified storage.
-
Configure data tiering policies
In the Hammerspace console, define policy rules to automatically move cold data from Tier 0 to cloud object storage or other tiers. Set thresholds based on access frequency or file age to optimize performance and capacity utilization.
Frequently Asked Questions
What is Hammerspace Tier 0 and how does it work?
Hammerspace Tier 0 is a software-defined storage solution that turns local NVMe drives inside GPU servers into a shared parallel file system. It uses standard NFS/pNFS protocols and Anvil metadata servers to coordinate I/O, eliminating the need for external flash arrays.
How does Hammerspace Tier 0 improve AI training performance?
Tier 0 reduces checkpointing time for AI training by up to 10x compared to external networked storage. It also delivers 2–2.5x higher throughput and 50% lower latency in cloud deployments, boosting GPU utilization and accelerating model iteration.
What is the locality affinitization feature in Hammerspace v5.2?
Locality affinitization automatically detects when a compute node hosts a local Tier 0 volume and prioritizes that volume in pNFS layouts. This ensures most I/O stays local, avoiding network round trips without manual directory configuration or proprietary kernel patches.
How does Hammerspace Tier 0 compare to Vast Data, Weka, and DDN with Lustre?
Unlike competitors that rely on external NVMe flash arrays and high-speed networking, Tier 0 repurposes existing server-local NVMe drives. This reduces hardware costs, power consumption, and rack space, but performance depends on local drives rather than a centralized flash pool.
What are the limitations of Hammerspace Tier 0?
Performance is limited by the number and speed of local NVMe drives per node. It requires Hammerspace Anvil metadata servers for coordination, and the affinitization feature is only available in v5.2+. Colder data tiers on object storage are still needed for capacity.
Can Hammerspace Tier 0 be deployed in the cloud and how does it perform?
Yes, Tier 0 can be deployed on-premises or in the cloud across all four major US cloud platforms. In cloud deployments, it delivers 2–2.5x higher throughput and 50% lower latency compared to native cloud file systems, reducing GPU idle time.
Alternatives
How Tier 0 compares
Direct head-to-head against 2 competitors. Picked by 7wData.
Tier 0
- Pricing
- Not specified in sources
- Target
- Hammerspace Tier 0 is a software-defined storage solution that transforms the local NVMe drives inside GPU servers into a high-performance, shared, parallel file system accessible
- Strength
- Reduces checkpointing time for AI training by up to 10x compared to external networked storage, enabling faster model iteration.
- Watch for
- Performance is inherently limited by the number and speed of local NVMe drives per node, rather than a centralized flash pool that can be scaled independently.
H&R Block
- Pricing
- $0-$85 for online filing, state additional
- Target
- Individual filers, from simple to self-employed
- Deployment
- Web and mobile
- Strength
- In-person tax office drop-off option nationwide
- Watch for
- Pricing can escalate with add-on services like audit support
TaxSlayer
- Pricing
- $0-$47 for federal, state additional
- Target
- Budget-conscious filers, freelancers, and small businesses
- Deployment
- Web and mobile
- Strength
- Lower base price for self-employed returns
- Watch for
- Less guided experience; fewer import options than TurboTax
User reviews
No user reviews yet. Be the first to write one.
Sources
Reporting on this tool draws on these publicly available sources.