Kioxia’s new XL-Flash SSD will deliver more than 10 million IOPS in small-block workloads. This is three times faster than the peak speed of today’s SSDs.
The company announced the development of at a corporate strategy meeting earlier last week. One of the performance bottlenecks of today’s AI servers is the transfer of data between the storage and the GPU using the CPU, which significantly increases latency. Kioxia is developing new controllerdesigned specifically for maximizing IOPS, which will ensure that graphics chips are always 100% utilized due to fast data access.
The proposed AI SSD will utilize single-level flash (SLC) XL-Flash memory, which has a read latency of only 3 to 5 microseconds. Modern drives based on conventional 3D NAND offer 40 to 100 microseconds. Also, the storage of one bit per cell in SLC provides shorter access times and greater endurance, which are critical for AI tasks.
Modern high-performance SSDs for the data center typically achieve 2 to 3 million IOPS in both 4KB random read operations and 512B. Large language models typically perform small random accesses to select data, parameters, or database records. In these scenarios, smaller blocks perform better, so it makes sense to use 512B blocks for LLM and RAG models.
Kioxia’s AI SSD will also be optimized for peer-to-peer communication between the GPU and the drive without the need for a CPU. GPUs typically operate with a self-consistent memory cache, so 512B reads are better aligned with their configurations. The Kioxia AI SSD is scheduled for release in the second half of 2026.
Source: Tom`s Hardware