ScaleFlux bets smarter SSD controllers can break the AI memory bottleneck
SiliconANGLE Jonathan Anthony
ScaleFlux says smarter SSD controllers can help AI chips stop waiting on data. That matters because GPUs are fast, but memory and storage are way slower.
Based on reporting by SiliconANGLE, Jonathan Anthony — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
ScaleFlux is betting that the real AI bottleneck has shifted away from raw compute and toward the messy business of moving data around. GPUs can run at about 20 to 22 terabytes per second, but once data drops into DRAM or storage, the slowdown is brutal. Pichai Balaji, the company’s vice president of products and marketing, put it bluntly: memory is roughly 37 to 40 times slower than what the GPU wants, and storage is around 800 times slower. That gap leaves expensive chips waiting around for the next batch of data.
The company’s answer is to make SSD controllers and Compute Express Link memory controllers do more of the thinking. ScaleFlux pairs those chips with firmware and a reference architecture, then aims them at key-value cache offload, the conversational memory that inference systems rely on. In other words, flash is no longer just a filing cabinet. It can act more like active memory for AI workloads if the controller is smart enough.
Balaji said ScaleFlux has shown more than 200 flexible data placement streams and 7 to 10 drive writes per day on standard triple-level-cell NAND. The pitch is practical as much as technical. ScaleFlux is NAND agnostic, so customers can build around whatever flash they can actually get in a tight supply market.
That matters because AI factories are being built under real hardware constraints, not glossy slides. If the data path stays clogged, the most expensive part of the system sits idle. ScaleFlux is trying to squeeze more useful work out of NAND instead of pretending the hierarchy below the GPU isn’t there.
My take — AI-written commentary, not fact-checked reporting
This is the rare AI hardware pitch that sounds like it was written by someone who has seen a memory chart before lunch. The industry keeps throwing more compute at the problem, then acts surprised when data movement turns into the bill. Smarter controllers are less sexy than bigger chips, which is exactly why they may matter more.
Read more about this at: SiliconANGLE