The latest cooperation between Kioxia and NVIDIA is regarded as another technological leap forward in AI data center infrastructure. The two companies are working together to develop SSDs that can provide up to 100 million IOPS (input and output operations per second). This performance is much higher than the current top SSD level of about 3 million IOPS, aiming to overcome the data transmission bottleneck in large-scale AI model training and deployment.
If the project is successful, it will not only accelerate the pace of generative AI (GenAI) innovation, but will also redefine the data center architecture, change the competitive landscape of cloud computing and enterprise-class storage, and set new standards for the entire hardware ecosystem.
Semiconductor memory manufacturer Kioxia is developing a new generation of SSD technology to meet the high-intensity reading and writing needs of AI. The company announced that it plans to achieve commercialization in 2027, so that the random IOPS of SSD will reach 100 million times. This performance indicator is approximately 30 to 35 times that of existing high-end products. Kioxia cooperates with NVIDIA this time to promote the implementation of the project.
At a media briefing in Tokyo, Kioxia said the new SSD will be directly connected to NVIDIA GPUs without going through a traditional server central processor. This "point-to-point" connection method greatly improves the speed of data flow between storage and computing units. It is especially suitable for large AI model training that relies on frequent, small-scale random data reading, such as retrieving embedded representations and model parameters. Traditional SSD systems are difficult to efficiently cope with these needs.
NVIDIA has set a more challenging goal: to achieve 200 million IOPS through two such SSDs with the support of the upcoming PCI Express 7.0 standard. PCIe 7.0 brings higher-speed point-to-point GPU communication capabilities. In comparison, the current high-performance SSD achieves about 3 million IOPS in 4K blocks. If it jumps to 100 million IOPS, it will bring major technical challenges to both NAND flash memory and interface architecture.
Kioxia's most promising technology is its proprietary XL-Flash single-layer cell (SLC) NAND storage, which has the characteristics of high endurance, low latency and strong performance. Each XL-Flash chip supports up to 16 "planes", while ordinary consumer-grade 3D NAND generally has 3 to 6 planes.
Although Kioxia has not released all technical specifications, the existing test data can give a glimpse of the scale of the challenge: a 400GB XL-Flash SSD equipped with 32 NAND chips and using the PCIe 5.0 interface has demonstrated a performance of approximately 3.5 million random read IOPS. Theoretically, if the performance can be perfectly linearly scaled, then an SSD equipped with 915 chips may achieve 100 million IOPS. However, in fact, due to factors such as controller bandwidth, firmware overhead, and system architecture, this goal is often difficult to achieve by simply stacking chips, and may require the use of multiple controllers or modular SSD solutions.
Given the limitations of traditional 3D NAND expansion, Kioxia is also exploring a new storage technology called high-bandwidth flash (HBF), which aims to combine the speed of high-bandwidth memory with greater storage capacity. The HBF solution uses advanced packaging to stack up to 16 NAND chips and one logic chip to achieve extremely high levels of parallelism and bandwidth. Although it is still uncertain whether HBF will be used in the final product of this cooperation project, this research and development shows that Kioxia is laying out ultra-high-performance storage solutions in the AI era.