The intersection of AI and storage infrastructure often
creates unexpected challenges. During this week’s IT Press Tour briefing,
Hammerspace showcased how they’re reshaping this landscape with their Tier 0
technology – and the implications are far more interesting than you might
expect.
“Two seismic shifts are fueling unprecedented growth – the rising need for cost and power-efficient infrastructure to support GPU computing at scale and the rapid adoption of hybrid cloud and multi-data center architectures,” said David Flynn, Hammerspace CEO and Co-founder.
The GPU Storage Challenge: A Hidden Problem in Plain
Sight
Here’s something that might surprise you: Most GPU servers
are sitting on a gold mine of unused storage.
“What we’re hearing from customers deploying large GPU
computing environments is that time-to-value is critical, as are infrastructure
cost and constraints,” explained Molly Presley, SVP Global Marketing at
Hammerspace.
“The reality is that GPU server local NVMe storage largely
goes unused because it’s siloed. Today, a GPU server may have 60-500TB of
storage, but later in 2025, 2PB of storage per server will be possible.”
That’s a lot of expensive storage capacity going to waste.
The problem stems from how GPU infrastructure has
traditionally been designed. Organizations typically deploy powerful GPU
servers packed with local NVMe storage, then add external storage systems for
data sharing and reliability. This creates a peculiar situation where:
- Local
storage remains trapped in silos, unused for its intended purpose - Additional
external storage systems consume extra power and rack space - Infrastructure
costs multiply unnecessarily - Deployment
times stretch longer than needed - Management
complexity increases with each new system
It’s a bit like buying a fleet of trucks with massive cargo
spaces but only using the driver’s cabin – while renting additional storage
space elsewhere.
The Tier 0 Breakthrough: Rethinking Storage Architecture
Hammerspace’s approach turns this problem on its head.
Instead of asking customers to buy more storage, they’ve developed a way to use
what’s already there. Their Tier 0 technology transforms that trapped local
storage into a shared resource that any GPU in the cluster can access.
The Technical Innovation Behind Tier 0
What makes this particularly interesting is the
implementation. Rather than requiring additional software installation on
compute nodes, Tier 0 works with standard Linux NFS capabilities. They’ve
cleverly leveraged the Linux 6.12 kernel’s local I/O patch, which enables the
system to recognize when it’s serving data to itself – creating a direct,
high-speed data path.
This means:
- No
additional software needed on compute nodes - Native
Linux kernel support - Direct
client-to-storage data paths - Peak
GPU performance without compromise
Real-World Impact: A Case Study That Changes Perspectives
Consider a real deployment scenario with 1,000 GPU servers
housing 8,000 GPUs total. In a traditional setup, organizations would have 30PB
of stranded local SSD storage essentially wasted, while simultaneously
investing in 20PB of external NVMe storage.
Hammerspace’s Tier 0 technology
unifies this into a single 50PB pool, eliminating the need for additional
hardware while delivering improved performance and lower power consumption.
The Complete Data Platform: Beyond Just Fast Storage
Hammerspace’s solution provides a comprehensive data
platform spanning from microsecond-access Tier 0 through to archive storage. At
the fastest end, Tier 0 delivers microsecond access through local NVMe, perfect
for active AI training data and checkpointing. Tier 1 provides millisecond
access via shared NVMe, supporting collaborative workflows and data
preparation. Traditional storage comprises Tier 2, offering cost-effective
capacity and wide-scale sharing, while archive storage rounds out the platform
with long-term retention capabilities.
Market Validation and Growth
The market has responded enthusiastically to Hammerspace’s
approach. Presley shared their impressive growth metrics: “We’ve achieved
10x year-over-year revenue growth, with a 32% increase in total customers. Our
gross revenue retention exceeds 95%, while our net revenue retention has
surpassed 330%.”
And the company has expanded internationally, establishing their
first customer wins in Germany, the Middle East, and India, while opening a new
office in Singapore to support their Asian market expansion.
Strategic Partnership with Hitachi Vantara
A collaboration with Hitachi Vantara represents a
significant milestone for Hammerspace. “We wanted to work with someone
like Hitachi because we know that enterprises are just starting to get into AI,
and they’re in 85% of all the Fortune 500,” Presley notes.
This partnership integrates Hammerspace’s innovative
technology with Hitachi’s enterprise reach, delivering comprehensive AI
infrastructure solutions with simplified deployment options and
enterprise-grade support.
Practical Benefits for IT Teams
For IT professionals managing AI infrastructure, the impact
of Hammerspace’s approach is substantial. Flynn emphasized the simplicity: “This is really remarkable – this whole
thing works without any Hammerspace software living on these nodes at all. You
orchestrate the data on the tier 0 as well, but you don’t have to deploy
anything.”
The platform dramatically improves GPU utilization through
faster data access while reducing power consumption and infrastructure costs.
Organizations can deploy the solution in under an hour, using standard Linux
protocols and a unified management interface that simplifies ongoing
operations.
Looking Ahead: The Future of AI Infrastructure
As AI workloads continue to grow in complexity, efficient
resource use becomes increasingly critical. “Storage is usually a much
narrower bottleneck,” Flynn explains. “By leveraging the existing
infrastructure and built-in Linux capabilities, we’re removing that bottleneck
without adding complexity.”
For organizations investing in AI infrastructure, this
approach could mean the difference between buying more external storage or
making better use of what they already have. As we move forward, solutions like
Hammerspace’s Tier 0 will likely become increasingly important in managing the
growing demands of AI workloads while keeping costs and complexity under
control. In today’s budget-conscious environment, that’s not just smart – it’s
necessary.
This shift in storage architecture represents a fundamental
change in AI infrastructure design, where efficiency and simplicity take
precedence over adding more hardware. As Flynn concludes, “It’s not just about
storage – it’s about transforming how we think about data access and management
in AI environments.”
##
