KIOXIA Europe GmbH has announced its new GP1 Series NVMe SSD model, designed to provide direct GPU access in AI infrastructures. Developed with PCIe 6.0 technology, this drive is the first member of the KIOXIA GP Series, which will be introduced in early 2026.

The new generation drive achieves 10 million random read IOPS at a block size of 512 bytes. This high performance is made possible by the company’s own developed controller and the SLC NAND memory architecture supported by second-generation XL-FLASH technology.

High Performance and Durability

Manufactured according to PCIe 6.0 and NVMe 2.2 standards, the GP1 Series stands out with its daily write (DWPD) durability rating of 50 drive writes. This high level of durability allows the driver to operate for extended periods without experiencing performance loss under demanding AI workloads.

The GP1 Series acts as a flash layer that expands high-bandwidth memory in AI systems. This process allows systems to access much larger datasets at low cost while also increasing GPU usage efficiency.

IOPS value refers to the number of read and write requests a storage device can process per second. A high IOPS value eliminates bottlenecks by enabling the rapid transfer of information from the storage unit to the processing unit in AI training and inference processes.

Calculated on a 3TB capacity model, a 50 DWPD value means approximately 109.5 PB of total write endurance over a two-year warranty period. This number demonstrates how robust the driver is for demanding enterprise-level operations.

Physical Design and Future Goals

The KIOXIA GP1 Series will be launched in E3.S and E1.S form factors, with 9.5 mm and 15 mm height options. The E3.S and 9.5 mm E1.S models offer liquid cooling support, while all configurations are compatible with standard air-cooled systems.

The company plans to deliver evaluation samples to select customers by the end of 2026. The widespread launch of the product will take place at the FMS conference in California.

The architecture of the GP1 Series is designed to be scalable, capable of reaching levels of 100 million IOPS in the future. This goal aligns with the performance levels that Nvidia demands from SSD manufacturers for AI infrastructure.

However, achieving these goals may require the development of entirely new memory categories beyond current SSD designs. Do you think a storage speed of 10 million IOPS is sufficient to provide the expected productivity increase in AI projects?