A PCIe 6.0 SSD built around GPU access

As of Saturday, 22 August 2026, the Kioxia GP1 Series is still a newly announced enterprise SSD family rather than a generally available product. Kioxia introduced it on 4 August 2026 at FMS: the Future of Memory and Storage in Santa Clara, positioning it as the first product in its GP Series of Super High IOPS SSDs optimized for GPU direct access. That detail matters more than the usual enterprise SSD pitch: GP1 is not mainly being framed as a bulk capacity drive or a sequential bandwidth showcase. It is aimed at AI infrastructure where GPUs need fast, repeated access to small chunks of data without waiting on slower storage paths. (kioxia.com)

Article contains affiliate links, commission may be earned.

The verified GP1 specs so far

Kioxia has confirmed several core details, while leaving other commercial specs such as capacity, pricing, and final mass-production configurations undisclosed. The GP1 Series is designed to the PCIe 6.0 and NVMe 2.2 specifications, uses second-generation Kioxia XL-FLASH low-latency flash memory, and targets up to 10 million random read IOPS using a 512-byte block size. Kioxia also lists up to 50 DWPD endurance. Form factors include E3.S and E1.S 9.5 mm / 15 mm, with all versions supporting air-cooled environments, while cold-plate liquid cooling is listed for E3.S and E1.S 9.5 mm configurations. (kioxia.com)

See Crucial T705 13 GB/s NVMe SSD price

Why 512-byte I/O is the interesting part

The headline 10 million IOPS figure is attention-grabbing, but the block size explains why Kioxia is steering the discussion toward AI systems. Large language model inference, retrieval pipelines, vector databases, and other accelerator-heavy workloads often involve many small, random reads rather than simple long sequential transfers. GP1’s 512-byte access target suggests a drive tuned for fine-grained data movement near the GPU, where feeding the accelerator quickly can matter as much as storing the data in the first place. Kioxia says the design enables flash-based memory extension and is intended to improve GPU utilization by making larger datasets accessible without relying only on in-node HBM or DRAM expansion. (americas.kioxia.com)

View NVIDIA DGX Spark Personal AI Desktop Supercomputer on partner website

A flash tier between HBM and normal storage

The bigger idea is a new memory hierarchy. HBM remains extremely fast and close to the GPU, but it is expensive and limited in capacity. Traditional SSDs offer much more capacity, but they are not usually treated as something a GPU can lean on like extended working memory. GP1 tries to occupy the space between those two worlds: not a replacement for HBM, and not just another back-end storage pool. Kioxia’s earlier GP Series material described this as high-speed flash memory that can expand usable GPU-accessible memory space for AI and high-performance computing workloads. (europe.kioxia.com)

Buy ASUS GeForce RTX 5090 here

What is still unknown in August 2026

For now, GP1 is best understood as a product direction and an early enterprise platform, not something buyers can compare like a retail SSD. Kioxia says evaluation samples will be available to select customers by the end of 2026, so real-world deployment details should become clearer after hyperscalers, AI labs, and server vendors test it in full systems. The company has also said the architecture is designed to scale from today’s 10 million random read IOPS toward future generations targeting 100 million IOPS, but those future figures should be treated as a roadmap goal, not a current GP1 specification. (kioxia.com)

See SANDISK Optimus GX PRO 8100 PCIe 5 SSD price