Kioxia and Nvidia have announced a joint project to raise storage performance for AI data center infrastructure. Their goal is an SSD capable of reaching 100 million IOPS, roughly 30 to 35 times higher than today's top-end drives, which top out at 3 million IOPS. If it succeeds, the breakthrough could not only shorten the time needed to train and deploy huge AI models but also reshape how data centers are designed, shifting the competitive landscape in cloud and enterprise storage.
A Next-Generation SSD Project for the AI Era
Kioxia, a leading maker of semiconductor memory, is developing a new SSD technology with extremely fast random reads to meet the demands of AI workloads. The company aims to commercialize an SSD that reaches 100 million IOPS in 2027, working directly with Nvidia on the project.
At a presentation in Tokyo, Kioxia said the new SSD will connect directly to Nvidia GPUs rather than going through the central processor (CPU). This peer-to-peer connection allows much faster data transfer between storage and compute resources, which is a key factor for large AI models that need constant random data access. Tasks such as fetching embeddings or model parameters will be optimized by the new mechanism.
Nvidia is aiming even higher: a system with two SSDs could reach 200 million IOPS using the PCI Express 7.0 standard, which supports high-speed communication between GPUs. For perspective, today's high-performance SSDs reach only about 3 million IOPS with 4K blocks. Jumping to 100 million IOPS will therefore demand major improvements in both NAND flash technology and the interface architecture.
XL-Flash and a New Technical Direction
One technology expected to underpin Kioxia's SSD is XL-Flash, a proprietary SLC NAND memory known for high endurance, low latency, and outstanding throughput. XL-Flash allows up to 16 planes in a single NAND die, while conventional 3D NAND has only 3 to 6 planes.
Although Kioxia has not released detailed specifications, test data shows the challenge is considerable. A 400GB XL-Flash SSD with 32 NAND dies and a PCIe 5.0 interface once reached about 3.5 million random read IOPS. With perfect scaling, a drive with 915 dies could hit 100 million IOPS. In practice, though, factors such as controller bandwidth, firmware latency, and architectural limits make ideal performance hard to reach. That may force manufacturers to adopt multi-controller designs or modular SSDs.
Alongside XL-Flash, Kioxia is also researching High-Bandwidth Flash (HBF), a new memory technology that combines the speed and scalability of HBM with far greater storage capacity. HBF can integrate up to 16 NAND chips with a logic die in a stacked structure, using advanced packaging techniques to maximize bandwidth and parallel processing. It is not yet clear whether HBF will appear in a final commercial product, but it signals that Kioxia is broadening its strategy and aiming for ultra-high-performance storage solutions for the AI era.
