The landscape of artificial intelligence is shifting from compute-intensive workloads to data-intensive ones, demanding a rethinking of how GPUs access memory. Kioxia has stepped forward with a solution: the Super High IOPS SSD series, designed to integrate directly with GPU architectures and extend high-bandwidth memory (HBM) capacity beyond its physical limits.
This isn't just an incremental upgrade—it's a fundamental reimagining of how AI systems handle data. Traditional HBM, while fast, is constrained by size and cost. The new Kioxia GP Series SSDs, leveraging the company's XL-FLASH Storage Class Memory, offer finer-grained data access (512 bytes) at lower power consumption per IO compared to conventional TLC SSDs. This makes them uniquely suited for GPU-initiated AI workloads, where every millisecond of latency and watt of power saved can translate to significant performance gains.
The collaboration with NVIDIA is part of the Storage-Next initiative, which aims to address the growing gap between compute power and memory capacity in AI systems. By enabling GPUs to access flash-based memory, Kioxia's SSDs effectively expand HBM capacity without the physical or thermal constraints that come with traditional memory modules. This shift could unlock larger datasets for AI models, improving GPU utilization by bringing more data closer to compute resources.
For PC builders and system designers, this represents a significant upgrade path. The ability to pair GPUs with high-performance, low-latency storage that operates at the 512-byte granularity level opens new possibilities for workload optimization. However, compatibility remains a critical consideration—systems will need to support this architecture, which may not be universal across existing hardware.
Kioxia has also announced another addition to its lineup: the CM9 Series PCIe 5.0 E3.S SSD, offering up to 25.6 TB of TLC capacity with 3 DWPD endurance. This drive is tailored for large-scale AI inference environments, where context windows and KV cache requirements are pushing the limits of traditional storage. Samples for both series are expected to begin shipping in Q3 2026.
What's confirmed: Kioxia's GP Series SSDs will be available as evaluation samples by year-end, with broader availability in Q3 2026. The CM9 Series is also set for Q3 2026 availability. What remains to be seen: the exact performance benchmarks and real-world efficiency gains these drives will deliver in AI workloads, as well as how widely this architecture will be adopted beyond NVIDIA's ecosystem.
