Kioxia has announced the GP1 Series, a high speed PCIe 6.0 NVMe SSD engineered specifically for direct GPU access in artificial intelligence clusters. The storage drive uses second generation XL FLASH memory to hit 10,000,000 random read IOPS while acting as an economical memory extension tier alongside high bandwidth memory. Evaluation samples will ship to select enterprise customers before the end of the year following its public demonstration at the FMS conference in Santa Clara.
Modern artificial intelligence workloads frequently stall when GPUs run out of high speed memory. Expanding high bandwidth memory or standard system DRAM inside server nodes is prohibitively expensive. Kioxia is positioning the GP1 Series as a direct solution to this bottleneck. By allowing graphics processors to access data directly from solid state media at granular 512 byte block sizes, server clusters can process massive datasets with reduced power consumption per transfer. The underlying architecture is engineered to scale even further, with future generations targeting up to 100,000,000 IOPS.
Neville Ichhaporia, senior vice president and general manager of the SSD business unit at Kioxia America, explained the role of the new hardware in data center design:
The AI memory wall is a critical challenge for our industry to solve on the way to growing the scale and capability of AI deployment. Kioxia XL FLASH is uniquely positioned to deliver the extremely high performance and low latency data access needed to support GPUs as a memory extension tier. It also offers a significantly lower cost per gigabyte than expanding in node HBM or DRAM capacity.
The hardware complies with the latest PCIe 6.0 and NVMe 2.2 specifications to ensure maximum bandwidth across enterprise server backplanes. It is built to handle heavy computational write cycles, offering endurance ratings up to 50 drive writes per day. Physical deployment options include E3.S as well as E1.S sizes in both 9.5 mm and 15 mm thicknesses. To manage thermal output in dense computing racks, the E3.S and 9.5 mm E1.S models support liquid cooling through dedicated cold plates, while all form factors remain compatible with standard air cooled server environments.
