It seems that the AI industry is now focusing on creating an inference-optimized AI SSD through NAND chips, as SK hynix plans to introduce a solution by 2027.
NVIDIA & SK hynix Are Co-Developing an "AI SSD" That Could Bring Massive Benefits to Inference Workloads
As traditional AI workloads shift from training to inference, there's a need to switch up the tech stack to ensure a low-latency, high-throughput environment. One of the reasons for this is how NVIDIA has decided to integrate general-purpose GDDR7 memory into the Rubin CPX GPU for prefill. According to a report by Chosun Biz, a similar approach is also expected for NAND chips. It is claimed that NVIDIA and SK hynix are developing a new SSD solution named "Storage Next" as an internal project, and it could revolutionize the NAND segment.
It is claimed that SK hynix plans to present a prototype by the end of next year, and that the AI SSD could scale up to a massive 100 million IOPS, which is significantly larger compared to what traditional enterprise SSDs feature. When examining the need for such a solution, it becomes clear that the current AI workload structure requires continuous access to massive model parameters, which cannot be accommodated by HBM or general-purpose DRAM products. The AI SSD would eventually allow a pseudo-memory layer, optimized for AI workloads.
The Korean giant is reportedly collaborating with NVIDIA on the Storage Next project, with the primary goal of enhancing throughput and energy efficiency through advanced NAND and controller architectures. While the goal certainly seems optimistic, it is also important to note that the NAND supply lines are already under massive pressure following demand for storage from CSPs and AI giants, and given that the AI SSD solution being discussed actually becomes mainstream, well, you could expect a DRAM-like situation with NAND Flash chips as well.
It appears that facilitating AI workloads comes at a high cost, which disrupts existing supply chains, leaving neither consumers nor suppliers sufficient time to react to the changing supply-demand situation. DRAM contract pricing is worsening with each day, and by the looks of it, NAND might be next.
About the author: Muhammad Zuhair is a hardware and technology reporter for Wccftech, specializing in the semiconductor industry and the complex interplay between technology, manufacturing, and geopolitics. His coverage focuses on the corporate strategies and technological roadmaps of industry giants like TSMC, NVIDIA, Samsung, and Intel. Zuhair's expertise lies in deconstructing complex topics such as fabrication nodes (e.g., 2nm process), the economic impact of policies like the CHIPS Act, and the strategic development of AI infrastructure from NVIDIA, AMD and Intel.
Follow Wccftech on Google to get more of our news coverage in your feeds.
