Solidigm Breaks Down Siloed Thinking About AI Infrastructure
No data center component is an island. While the industry conversation around AI infrastructure has focused heavily on graphics processing units (GPUs), a more fundamental truth is emerging: peak performance requires optimizing every component in the stack to work together. In fact, attempting to optimize any single element without considering the broader system inevitably limits what’s possible.
I recently had the opportunity to explore this interconnected reality with Solidigm’s Ace Stryker, product marketing director of AI infrastructure, and Jeniece Wnorowski, director of industry expert programs, to understand how storage requirements are evolving. During our conversation, it became clear that the most significant breakthroughs in AI infrastructure are coming from rethinking how data flows through the entire AI pipeline.
AI’s Expanding Diversity Shapes Storage Requirements
Ace started our conversation by emphasizing that we are still only a few years into AI becoming a prominent cultural force, and that its diverse potential is still yet to be fully uncovered. As AI-enabled workloads diversify with new models, tools, and solution stacks, the requirements for hardware—including storage—are diversifying as well. Truly understanding these requirements, however, means looking beyond storage specifications alone. “From a storage perspective, we’re really concerned about how the storage in an AI cluster interacts with the memory to deliver optimized outcomes,” Ace said. “Storage does not do the job on its own.”
The challenge extends beyond simple read-write speeds. Modern AI systems require careful orchestration between storage layers, host dynamic random-access memory (DRAM), and high-bandwidth memory on GPUs. Understanding how data moves between these memory tiers has become essential for IT architects planning next-generation infrastructure. As Ace noted, attempting to optimize storage in isolation limits the ability to understand what’s actually happening in the AI pipeline.
Challenging Conventional Wisdom
This focus on the interaction between memory and storage has led to research with surprising outcomes. Solidigm recently worked with Metrum AI to examine what happens when significant amounts of AI data are strategically moved from memory onto solid state drives (SSDs) in ways that weren’t typically considered.
The companies used video from a busy traffic intersection and fed it into an analysis pipeline that generated embeddings and created a RAG database, then created a report about what happened in the video with suggestions for safety improvements. By offloading RAG data and inactive model weights from memory to SSDs, they achieved a 57% reduction in DRAM usage for a 100 million vector dataset. More surprisingly, queries per second actually increased by 50% compared to keeping data in memory, thanks to more efficient indexing algorithms in the SSD offload approach.
The implications extend beyond cost savings. The research demonstrated running the Llama 3.3 70 billion parameter model on an NVIDIA L40S GPU, a combination that normally exceeds the GPU’s memory constraints. For organizations looking to repurpose legacy hardware or deploy AI capabilities in edge environments with power limitations, this represents new possibilities for using hardware previously considered inadequate for modern AI-enhanced workloads.
The Density Revolution
While performance optimization captures headlines, capacity evolution tells an equally compelling story. Solidigm’s 122 terabyte (TB) drives, roughly the size of a deck of cards, represent just one milestone in a rapid progression that’s seen capacities jump from 30TB to 60TB to 122TB in a single year. The company has announced plans for 256TB drives, and as Ace said, “It’s not too long before you’re going to see Solidigm and others aiming at a petabyte in a single device, which was unfathomable even five years ago.”
These density improvements deliver practical benefits across the infrastructure stack. Higher capacity per drive means fewer physical devices required, reducing rack space requirements, power consumption, and cooling costs while maintaining the throughput AI-enabled workloads demand.
The Partnership Imperative
Throughout the conversation, Ace returned repeatedly to collaboration as Solidigm’s core operating principle. The company’s logo, an interlocking “S” design, symbolizes partnerships fitting together to solve complex problems. It’s reflected in their approach across the ecosystem, from working with NVIDIA on thermal solutions to collaborating with software orchestration leaders and cloud service providers.
This partnership focus acknowledges a fundamental reality: storage optimization happens within a broader system context involving networking, software orchestration, and compute resources. Solutions that work in isolation rarely deliver optimal outcomes at scale.
The TechArena Take
As AI-enhanced workloads continue their exponential growth, the organizations that understand storage as a strategic enabler rather than a commodity component will gain sustainable advantages. Solidigm’s research demonstrates that intelligent storage strategies can unlock performance improvements while simultaneously reducing costs and expanding deployment possibilities. For IT architects planning next-generation AI infrastructure, the message is clear: look beyond the GPU specifications and examine how data moves through your entire system. The gains not only in efficiency, but overall capability, may surprise you.
Learn more about Solidigm’s AI-focused storage innovations at solidigm.com or connect with Ace Stryker on LinkedIn.
No data center component is an island. While the industry conversation around AI infrastructure has focused heavily on graphics processing units (GPUs), a more fundamental truth is emerging: peak performance requires optimizing every component in the stack to work together. In fact, attempting to optimize any single element without considering the broader system inevitably limits what’s possible.
I recently had the opportunity to explore this interconnected reality with Solidigm’s Ace Stryker, product marketing director of AI infrastructure, and Jeniece Wnorowski, director of industry expert programs, to understand how storage requirements are evolving. During our conversation, it became clear that the most significant breakthroughs in AI infrastructure are coming from rethinking how data flows through the entire AI pipeline.
AI’s Expanding Diversity Shapes Storage Requirements
Ace started our conversation by emphasizing that we are still only a few years into AI becoming a prominent cultural force, and that its diverse potential is still yet to be fully uncovered. As AI-enabled workloads diversify with new models, tools, and solution stacks, the requirements for hardware—including storage—are diversifying as well. Truly understanding these requirements, however, means looking beyond storage specifications alone. “From a storage perspective, we’re really concerned about how the storage in an AI cluster interacts with the memory to deliver optimized outcomes,” Ace said. “Storage does not do the job on its own.”
The challenge extends beyond simple read-write speeds. Modern AI systems require careful orchestration between storage layers, host dynamic random-access memory (DRAM), and high-bandwidth memory on GPUs. Understanding how data moves between these memory tiers has become essential for IT architects planning next-generation infrastructure. As Ace noted, attempting to optimize storage in isolation limits the ability to understand what’s actually happening in the AI pipeline.
Challenging Conventional Wisdom
This focus on the interaction between memory and storage has led to research with surprising outcomes. Solidigm recently worked with Metrum AI to examine what happens when significant amounts of AI data are strategically moved from memory onto solid state drives (SSDs) in ways that weren’t typically considered.
The companies used video from a busy traffic intersection and fed it into an analysis pipeline that generated embeddings and created a RAG database, then created a report about what happened in the video with suggestions for safety improvements. By offloading RAG data and inactive model weights from memory to SSDs, they achieved a 57% reduction in DRAM usage for a 100 million vector dataset. More surprisingly, queries per second actually increased by 50% compared to keeping data in memory, thanks to more efficient indexing algorithms in the SSD offload approach.
The implications extend beyond cost savings. The research demonstrated running the Llama 3.3 70 billion parameter model on an NVIDIA L40S GPU, a combination that normally exceeds the GPU’s memory constraints. For organizations looking to repurpose legacy hardware or deploy AI capabilities in edge environments with power limitations, this represents new possibilities for using hardware previously considered inadequate for modern AI-enhanced workloads.
The Density Revolution
While performance optimization captures headlines, capacity evolution tells an equally compelling story. Solidigm’s 122 terabyte (TB) drives, roughly the size of a deck of cards, represent just one milestone in a rapid progression that’s seen capacities jump from 30TB to 60TB to 122TB in a single year. The company has announced plans for 256TB drives, and as Ace said, “It’s not too long before you’re going to see Solidigm and others aiming at a petabyte in a single device, which was unfathomable even five years ago.”
These density improvements deliver practical benefits across the infrastructure stack. Higher capacity per drive means fewer physical devices required, reducing rack space requirements, power consumption, and cooling costs while maintaining the throughput AI-enabled workloads demand.
The Partnership Imperative
Throughout the conversation, Ace returned repeatedly to collaboration as Solidigm’s core operating principle. The company’s logo, an interlocking “S” design, symbolizes partnerships fitting together to solve complex problems. It’s reflected in their approach across the ecosystem, from working with NVIDIA on thermal solutions to collaborating with software orchestration leaders and cloud service providers.
This partnership focus acknowledges a fundamental reality: storage optimization happens within a broader system context involving networking, software orchestration, and compute resources. Solutions that work in isolation rarely deliver optimal outcomes at scale.
The TechArena Take
As AI-enhanced workloads continue their exponential growth, the organizations that understand storage as a strategic enabler rather than a commodity component will gain sustainable advantages. Solidigm’s research demonstrates that intelligent storage strategies can unlock performance improvements while simultaneously reducing costs and expanding deployment possibilities. For IT architects planning next-generation AI infrastructure, the message is clear: look beyond the GPU specifications and examine how data moves through your entire system. The gains not only in efficiency, but overall capability, may surprise you.
Learn more about Solidigm’s AI-focused storage innovations at solidigm.com or connect with Ace Stryker on LinkedIn.
.png)
