Kioxia's GP1 SSDs hit 10 million IOPS to extend GPU memory with flash

Kioxia's GP1 Series reaches 10 million random read IOPS at 512-byte blocks on PCIe 6.0 with second-generation XL-FLASH, at up to 50 drive writes per day of endurance. Evaluation samples reach select customers only by the end of 2026, which makes this a roadmap stake rather than a shipping product. The play is flash as a cheaper tier beneath High Bandwidth Memory for feeding GPUs, an architecture Kioxia designed to scale toward 100 million IOPS.

Source: kioxia.com ↗

This approach lets AI systems access larger datasets at a significantly lower cost than adding more HBM, while improving GPU utilization.

Kioxia Corporation

Why this matters

→ Cuts AI memory costs by layering cheap flash beneath expensive HBM
→ 10M IOPS on PCIe 6.0 unlocks GPU-direct storage as a viable tier
→ Roadmap to 100M IOPS signals flash as structural GPU memory extension

Flash as GPU memory