Breaking the bottleneck: Why AI demands an SSD-first future
Presented by Solidigm
As AI adoption increases, data centers face a critical storage bottleneck – and traditional HDDs are at the heart of it. Data that once lay dormant as cold archives is now widely used to build more accurate models and deliver better inference results. This shift from cold data to warm data requires low-latency, high-throughput storage that can handle parallel computations. HDDs will remain the workhorse for low-cost cold storage, but without rethinking their role, the high-capacity storage layer risks becoming the weakest link in the AI factory.
“Modern AI workloads, combined with data center constraints, have created new challenges for HDDs,” said Jeff Janukowicz, research vice president at IDC. “As HDD vendors address the growth of data storage by offering larger drives, this often comes at the cost of slower performance. As a result, the concept of ‘nearline SSDs’ is becoming an increasingly relevant topic of discussion within the industry.”
Today, AI operators must maximize GPU utilization, efficiently manage network-attached storage, and scale computing power – all while reducing the cost of increasingly scarce energy and space. In an environment where every watt and every square centimeter counts, says Roger Corell, senior director of AI and leadership marketing at Solidigm, success requires more than a technical innovation. It calls for a deeper realignment.
“It speaks to the tectonic shift in the value of data for AI,” says Corell. “That’s where high-capacity SSDs come into play. Together with capacity, they deliver performance and efficiency – allowing exabyte-scale storage pipelines to keep pace with the relentless pace of dataset size. All of that consumes power and space, so we need to do it as efficiently as possible to enable more GPU scale in this constrained environment.”
High-capacity SSDs don’t just replace HDDs; they also remove one of the biggest bottlenecks on the AI factory floor. Delivering massive gains in performance, efficiency, and density, SSDs free up the power and space needed to push GPU scale further. It’s not so much a storage upgrade as it is a structural change in the way data infrastructure is designed for the AI era.
HDDs vs. SDDs: more than just a hardware update
HDDs have an impressive mechanical design, but they consist of many moving parts that consume more energy, take up more space, and fail faster than SSDs. The reliance on spinning platters and mechanical read/write heads inherently limits Input/Output Operations Per Second (IOPS), creating bottlenecks for AI workloads that require low latency, high concurrency, and sustained throughput.
HDDs also struggle with latency-sensitive tasks, because the physical search for data introduces mechanical delays that are not suitable for real-time AI inference and training. Additionally, their power and cooling needs increase significantly with frequent and intensive data access, reducing efficiency as data scales and heats up.
In contrast, the SSD-based VAST storage solution reduces energy consumption by ~$1M per year, and in an AI environment where every watt matters, this is a huge benefit for SSDs. To demonstrate this, Solidigm and VAST Data completed a study on the economics of data storage at exabyte scale – a quadrillion bytes, or a billion gigabytes, with an analysis of the energy consumption of storage versus HDDs over a ten-year period.
As a starting point, you’ll need four 30TB hard drives to match the capacity of a single 122TB Solidigm SSD. After taking into account VAST’s data reduction techniques, enabled by the superior performance of SSDs, the exabyte solution includes 3,738 Solidigm SSDs versus more than 40,000 high-capacity HDDs. The study found that the SSD-based VAST solution consumes 77% less storage energy.
Minimizing the footprint of data centers
“We ship 122 terabyte drives to some of the best OEMs and leading AI cloud service providers in the world,” says Corell. “If you compare a full 122TB SSD to a hybrid HDD + TLC SSD configuration, they get a nine to one savings in the data center footprint. And yes, it’s important in these huge data centers that build their own nuclear reactors and sign hefty power purchase agreements with renewable energy suppliers, but it becomes increasingly important as you get to the regional data centers, the local data centers and any other way out to your edge deployments where space may be at a premium.”
These nine-to-one savings go beyond space and power: organizations can fit infrastructure into previously unavailable spaces, expand GPU scale, or build smaller footprints.
“If you’re given percent fewer storage spaces to maintain and the associated costs. that’s gone.”
Another often overlooked element: the (much) larger physical footprint of data stored on mechanical hard drives results in a larger footprint of building materials. Together, concrete and steel production are responsible for more than 15% of global greenhouse gas emissions. By reducing the physical footprint of storage, high-capacity SSDs can help reduce concrete and steel emissions by more than 80% compared to HDDs. And in the final phase of the sustainability life cycle, namely the end of the drive life, there will be 90% fewer drives that need to be disposed of. .
Reshaping cold storage and archiving strategies
Moving to SDD isn’t just a storage upgrade; it is a fundamental realignment of data infrastructure strategy in the AI era, and it is accelerating.
“Large hyperscalers are trying to get the most out of their existing infrastructure, doing unnatural actions, if you will, with HDDs like overprovisioning them to almost 90% to try to wring out as much IOPS per terabyte as possible, but they’re starting to get by,” says Corell. “Once they move to modern, high-capacity storage infrastructure, the industry as a whole will follow that trajectory. Additionally, we’re starting to see these lessons about the value of modern storage in AI being applied to other segments such as big data analytics, HPC, and more.”
While all-flash solutions are almost universally embraced, there will always be a place for HDDs, he adds. HDDs will continue to exist in applications such as archiving, cold storage, and scenarios where the pure cost per gigabyte outweighs the need for real-time access. But as the token economy heats up and companies realize value in data monetization, the warm and warming data segments will continue to grow.
Solving the energy challenges of the future
Now in its 4th generation, with more than 122 cumulative exabytes shipped to date, Solidigm’s QLC (Quad-Level Cell) technology has led the industry in balancing higher drive capacities with cost-efficiency.
“We don’t think of storage as just storing bits and bytes. We think about how we can develop these amazing drives that can deliver benefits at the solution level,” says Corell. “The shining star here is our recently launched E1.S, specifically designed for dense and efficient storage in direct-attached storage configurations for the next generation fanless GPU server.”
The Solidigm D7-PS1010 E1.S is a breakthrough, the industry’s first eSSD with single-sided direct-to-chip liquid cooling technology. Solidigm worked with NVIDIA to address the dual challenges of thermal management and cost efficiency, while delivering the high performance needed for demanding AI workloads.
“We are rapidly moving to an environment where all critical IT components will be liquid-cooled directly on the chip on the direct-attach side,” he says. “I think the market needs to look at their approach to cooling because energy constraints and energy challenges are not going to diminish in my lifetime at least. They need to apply a neo-cloud mindset to how they design the most efficient infrastructure.”
Increasingly complex inferences push against a memory wall, making storage architecture a primary design challenge and not an afterthought. High-capacity SSDs, combined with liquid cooling and efficient design, are emerging as the only way to meet the escalating demands of AI. The mandate now is to build an infrastructure not just for efficiency, but for storage that can scale efficiently as data grows. The organizations that realign storage now will be able to scale AI tomorrow.
Sponsored articles are content produced by a company that pays for the post or has a business relationship with VentureBeat, and is always clearly marked. For more information please contact sales@venturebeat.com.
