Kioxia announced the GP1 Series PCIe 6.0 NVMe SSD on August 3, claiming up to 10 million random-read IOPS at a 512-byte block size using second-generation XL-FLASH. Evaluation samples are due to select customers by the end of 2026, so the headline performance remains a vendor specification rather than an independently validated production result.
Kioxia announced the GP1 Series on August 3, introducing the first named product in its high-IOPS GP family for GPU-accessible AI storage. The PCIe 6.0 NVMe 2.2 drive uses second-generation XL-FLASH and is specified for up to 10 million random-read IOPS at a 512-byte block size.
A fast flash tier for AI systems
Kioxia positions GP1 as a memory-extension tier for workloads whose active data exceeds available high-bandwidth memory. Instead of treating the drive as ordinary bulk storage, the design emphasizes small, low-latency reads that can keep accelerators supplied with data while moving less frequently used material out of HBM or DRAM.
That distinction matters when interpreting the performance claim. The 10 million figure is an “up to” random-read result using 512-byte blocks, not a general throughput guarantee or a direct comparison with the 4 KB measurements commonly published for enterprise SSDs. Kioxia says actual performance varies with the host, software and workload.
Hardware details and availability
The GP1 is designed for PCIe 6.0 and NVMe 2.2 and will be offered in E3.S plus 9.5 mm and 15 mm E1.S form factors. Kioxia lists cold-plate liquid-cooling support for the E3.S and 9.5 mm E1.S versions, air cooling across all versions, and endurance of up to 50 drive writes per day.
Kioxia plans to provide evaluation samples to selected customers by the end of 2026 and says those units are for functional checks; mass-production specifications may differ. The company also plans to show the drive at FMS in Santa Clara from August 4 through August 6.
What remains to be proven
The announcement does not disclose capacities, price, random-write performance, sequential bandwidth or measured end-to-end gains for a production AI workload. It also does not provide independent benchmark results.
Infrastructure teams should therefore treat GP1 as a concrete product milestone, not yet a purchasing baseline. Useful evaluation will require workload-specific latency distributions, power and cooling data, capacity economics, software-path compatibility and comparisons against additional HBM, DRAM and conventional NVMe storage.
Key Points #
- 1Kioxia announced GP1 on August 3 as the first named product in its high-IOPS GP Series for GPU-accessible AI storage.
- 2The company specifies up to 10 million random-read IOPS at a 512-byte block size using second-generation XL-FLASH.
- 3Evaluation samples are planned for select customers by the end of 2026, and Kioxia says production specifications may differ.
Scoring Rationale #
GP1 makes Kioxia’s GPU-accessible flash tier a concrete PCIe 6.0 product with an unusually high stated random-read ceiling, but samples are not due until late 2026 and key capacity, price and independent performance evidence remain unavailable.
Sources #
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.