January saw SK Hynix spin out a new venture tentatively called AI Company (AI Co.), with a $10 billion investment in its U.S.-based Solidigm arm. Originally a non-volatile (NAND) and solid state drive (SSD) business owned by Intel, Solidigm was acquired by Hynix for $9 billion in 2020 and merged with the South Korean giant’s own NAND operations.

As part of its AI Co. roadmap, SK Hynix completed its business restructuring of Solidigm around the same time as this year’s Mobile World Congress (MWC) event, with the assets and personnel of its NAND business transferred to the newly founded Solidigm entity.

While memory bottlenecks were not a central theme at this month’s MWC, SDxCentral did hear directly from operator SK Telecom (SKT) that it sees memory bottlenecks in AI workloads as easily resolvable thanks to Hynix’s high-bandwidth memory (HBM) technology, which is leveraged across the SK Group. The comment came as the operator pivots into AI data center operations, further bolstering the parent group’s footprint in the AI race.

Where Solidigm fits into the memory narrative is with its approach to bits-per-cell scaling, according to Roger Corell, senior director for AI and strategic marketing at Solidigm.

“The No. 1 challenge is supply," Corell said in an interview after this year's MWC event. "So you try to get as much as you can out of a given wafer, and scaling bits-per-cell gets you to more output. Scaling is the fastest way to get there, and we are a leader in bits-per-cell, which has manifested where we are now with the industry's highest capacity drive in high volume production."

When gates fly

Solidigm's “bet,” as Corell put it, was placed on the floating gate method for NAND flash cells to store electrical charge to represent data in SSDs. This is as opposed to the charge trap method.

In floating gate designs, each cell uses a tiny, well‑isolated conductive “floating gate” to hold electrons, which helps keep each cell’s charge clearly separated from its neighbors and can be advantageous when pushing to higher bits-per-cell at large capacities. In charge trap designs, electrons are stored in a special insulating layer that “traps” charge in many localized sites. This structure, which is widely used by many vendors, can be easier to manufacture in more loaded 3D NAND stacks.

“This is not to say that charge trap is not very good at things, at certain storage workloads,” Corell said. “But when we're talking in a world that needs increasingly massive amount of storage, that kind of distinct cell isolation prevents charge spread, which can disrupt adjacent cells.”

Indeed, the Hynix assimilation gave Solidigm access to both technologies, a USP in Corell’s book, with Solidigm the “only pure play” enterprise SSD company that isn’t “distracted by other lines of business.”

“We're No. 2 in the market now in terms of enterprise SSDs, as measured by market segment share. And we believe that … we've got line of sight to be No. 2 in enterprise SSDs, that is our objective," Corell said.

Roadmap ahead

With the full force of Hynix’s restructure behind it, and a global workforce more than 2,000 strong, Solidigm believes it can become the leading storage provider for AI.

“And when I say AI, I don't mean just AI data centers. I mean AI at the edge as well,” Corell said, pointing to AI growth as stemming from inference at the edge.

“Inference capacity requirements are pushing well past the memory wall,” Corell added, referring to a growing performance gap between fast processors and relatively slow memory, leaving a lag on system performance as processors spend significant time waiting for data.

“You can't do it with gigabytes of storage. You need terabytes. You need tens-of-terabytes of storage for inference,” Corell continued.

With a crystal ball in front of him, Corell promised Solidigm’s floating gate roadmap would strengthen, with upcoming capabilities deepening the firm's ecosystem.

EDSFF future

The director was more to the point when discussing enterprise and data center SSD form factor (EDSFF), a family of newer SSD shapes and connector standards designed specifically for servers and data centers. Solidigm sees EDSFF as the natural SSD companion to AI, packing more flash into a server while maintaining decent airflow and power delivery, and designed for maximum rack efficiency.

“The first EDSFF spec came out in 2018 or 2017, maybe earlier. And the industry is every year like, 'this is the moment for EDSFF,'" Corell said. "Well, let me tell you AI is it. AI is the moment. When you look at the density requirements, the necessity to just have every watt and every square inch matter in the data center, in AI data centers, and how EDSFF is the far superior technology for that. So we're going to continue to work on that portfolio.”

One aspect of Solidigm's future that can't yet be revealed is its hold as the lynchpin of Hynix’s AI Co. ambitions. Corell stressed Solidigm will continue to stand on its own two feet, as opposed to being the third part of a hypothetical SKT/Hynix-led triad.

“We've got the backing of a massive corporation. … [But] Hynix has a unique set of customer base, and we have a customer base. And so I think if you look at it kind of simplistically, from customer base, there's very clear delineation,” Corell said.