Kioxia and Nvidia develop SSDs 100x faster for next-gen servers
Chiang, Jen-Chieh, Taipei; Sherri Wang, DIGITIMES Asia
Sunday 5 October 2025
Kioxia, Japan's leading memory chip maker, is partnering with Nvidia to develop a revolutionary solid-state drive (SSD) nearly 100 times faster than conventional models. The new drive, designed for generative AI servers, is scheduled for a 2027 debut...
Copilot expansion:
Kioxia and NVIDIA are co-developing ultra-fast SSDs for AI workloads that deliver up to 100 million IOPS—nearly 100× faster than today’s SSDs—by directly connecting to GPUs and bypassing CPUs. These SSDs are expected to launch in 2027 and support PCIe 7.0.
Here’s the deeper breakdown:
What Makes These SSDs 100× Faster?
- IOPS Target: Kioxia is building SSDs that hit 100 million IOPS (Input/Output Operations Per Second)—a massive leap from today’s top-tier SSDs, which typically max out around 1 million IOPS.
- GPU-Direct Architecture: Instead of routing data through CPUs, these SSDs will connect directly to NVIDIA GPUs, drastically reducing latency and boosting throughput for AI inference and training.
- PCIe 7.0 Interface: They’ll support PCIe Gen 7, which offers up to 512 Gbps of bandwidth—4× faster than PCIe 5.0—enabling ultra-low latency and high-speed data exchange.
Why This Matters for AI
- HBM Alternative: NVIDIA is exploring these SSDs as partial replacements for HBM (High Bandwidth Memory), expanding memory pools without relying solely on expensive DRAM.
- Memory Expansion Strategy: With two SSDs per GPU, NVIDIA aims to hit 200 million IOPS, enabling larger model contexts and faster inference cycles.
- AI-Driven NAND Demand: Kioxia projects that AI workloads will drive nearly half of NAND demand by 2029, reshaping the flash memory market.
Under the Hood: XL-Flash and NAND Tech
- Kioxia may leverage Gen 2 XL-Flash in SLC mode (1 bit/cell), which offers:
- <5 µs read latency
- Up to 14 GB/s sequential read
- 3.5 million IOPS per SSD in current prototypes
- To reach 100M IOPS, Kioxia will likely scale plane-level parallelism, controller architecture, and NAND die density.
Timeline and Market Impact
- Sample Shipments: Expected in late 2026
- Commercial Launch: Targeted for 2027
- Market Disruption: Could trigger a NAND shortage by 2026–2027 as AI clusters shift from HDDs and DRAM to flash-based memory expansion
This move could redefine GPU memory architecture, especially for GenAI and LLM workloads. If you're benchmarking XL-Flash vs HBM4 or mapping PCIe 7.0 rollout across AMD/NVIDIA platforms, I can help chart that next. Want to dive into NAND vs HBM economics or SSD controller architecture? |
