Shaping a New Form of Storage with BiCS Flash and XL-Flash

With Kioxia setting 2028 as the target launch date for its 100 million IOPS SSD, the Kioxia GP Series, built on XL-Flash, is opening up entirely new possibilities for AI computing. Looking at the present, enterprise- and data centre-class SSDs based on the 9th and 10th generations of BiCS Flash are also poised for launch, whilst PCIe 6.0 SSD products are gradually coming to the attention of enterprise users. The storage sector is rapidly evolving in response to the surging demand for AI and has already established entirely new storage solutions. It is evident that flash memory is playing an increasingly vital role in the AI era.

Achieving decoupling of storage and logic through CBA technology

Both the 9th and 10th generation BiCS Flash are based on the CBA (CMOS directly Bonded to Array) architecture established by the 8th generation BiCS Flash. This involves manufacturing the CMOS peripheral logic circuits and the memory cell array on two separate wafers, which are then bonded together using high-precision wafer-to-wafer alignment. This decoupling allows the memory array and logic units to evolve and enter production independently.

CBA technology provides the technical foundation for the ninth-generation BiCS Flash to achieve high performance at a lower investment cost, and for the tenth-generation BiCS Flash to surpass ultra-high capacity and deliver even higher performance through advanced stacking. The ninth-generation BiCS Flash will primarily target markets sensitive to production capacity and reliability, such as mobile devices, client SSDs and UFS. By reusing mature processes for storage arrays and upgrading the CMOS ‘brain’, upgrades can be implemented on existing production lines. Currently, both the 512Gb TLC and 1Tb TLC based on the ninth-generation BiCS Flash are poised for launch.

In terms of performance, compared with the sixth-generation products of the same capacity, the 512Gb TLC based on the ninth-generation BiCS Flash delivers a 61 per cent improvement in write performance, a 12 per cent improvement in read performance, a 36 per cent improvement in write energy efficiency, and a 27 per cent improvement in read energy efficiency. The Toggle DDR 6.0 interface achieves a data rate of 3.6 Gb/s, with a maximum of 4.8 Gb/s under demonstration conditions.

The 10th-generation BiCS Flash utilises higher costs to achieve performance and high storage density; it is primarily aimed at enterprise- and data centre-class SSDs, targeting the demands of AI storage for high performance, large capacity and low power consumption. Kioxia began supplying samples of its 1Tb TLC product based on the 10th-generation BiCS Flash in July this year; this product is manufactured at Fab 2 in the Kitakami plant in Iwate Prefecture.

The 10th-generation BiCS Flash features a 332-layer stack and optimised lateral density, delivering a 59 per cent increase in bit density compared to the 8th generation, whilst its NAND interface speed of 4.8 Gb/s represents a 33 per cent improvement over the 8th-generation BiCS Flash. Furthermore, the corresponding QLC version has pushed bit density to >37 Gb/mm², representing a 60 per cent increase over the eighth generation, and has become the world’s first QLC 3D flash memory to achieve a 4.8 Gb/s interface speed.

Accelerating AI Data Flows

XL-Flash is a storage-class memory (SCM) built on BiCS Flash, designed to combine the low latency of memory with the non-volatility of flash, thereby bridging the long-standing performance gap between DRAM and SSDs. The first-generation XL-Flash featured 16 planes and 4KB pages, with read latency of less than 5 μs, making it a competitor to Optane at the time. With the second-generation XL-Flash now in mass production and the addition of an MLC variant, the die capacity has doubled to 256 Gb, enabling significant cost reductions.

The forthcoming third-generation XL-Flash will utilise CBA technology to reduce latency to less than 10 per cent of that of BiCS Flash. Designed specifically for ultra-high IOPS, it will become a key component in unlocking AI computing power in the future.

Currently, Kioxia’s GP series, which utilises second-generation XL-Flash in conjunction with the PCIe 6.0 protocol, has already achieved a performance of 10 million IOPS. New products based on third-generation XL-Flash and paired with PCIe 7.0 will directly target 100 million IOPS. These new products will reshape the latency and tiering landscape of storage, providing robust support for generative AI workloads that are extremely sensitive to latency and IOPS, such as LLM training datasets and RAG vector stores.

It is evident that the CBA architecture is helping BiCS Flash—and indeed XL-Flash—to bring about a new era of transformation across the entire storage ecosystem. By continuously advancing its ultra-high-speed IOPS SSD products, Kioxia is providing robust support for the lower-tier cache layers of HBM.

As capacity and cost bottlenecks in GPU-side HBM become increasingly apparent, the offloading of hot data—such as KV Cache and vector indexes—is becoming a critical requirement for AI training and inference. Meanwhile, XL-Flash is precisely filling the long-standing performance gap between HBM/DRAM and traditional NAND with microsecond-level latency—comparable to that of memory—and non-volatile, high-capacity storage. From BiCS Flash’s reimagining of capacity and cost to XL-Flash’s reorganisation of the storage pyramid, Kioxia is transforming storage from a passive data container into an active component that coordinates with computing power and directly participates in AI data flows.