Kioxia Eyes GPU Memory Stack with High-Bandwidth Flash Push

Kioxia Eyes GPU Memory Stack with High-Bandwidth Flash Push

Blocks & Files
Blocks & FilesMar 20, 2026

Why It Matters

The technologies promise to alleviate storage‑to‑GPU bottlenecks, enabling cheaper, higher‑performance AI training and inference at scale. They also position Kioxia as a primary supplier for next‑generation data‑center memory architectures.

Key Takeaways

  • GP Series targets GPU servers with 10M IOPS by 2026
  • XL‑Flash built on 218‑layer Gen 8 technology
  • HBF stacks 32 dies via TSVs, near GPU bus
  • Optical SSD prototype transmits PCIe signals 100 m optically
  • Kioxia‑SanDisk JV extended past 2030; WD refocuses on HDDs

Pulse Analysis

Kioxia’s newly announced GP Series aims to close the bandwidth gap between storage and graphics processing units in data‑center servers. Built on the company’s Gen 2 XL‑Flash technology, the drive uses a 218‑layer, generation‑8 NAND stack and is housed in an E1.S PCIe 5 enclosure, with an E3.S version planned. Kioxia projects the GP Series to deliver 10 million IOPS by 2026 and scale to 100 million random‑read IOPS by 2027, offering a cost‑effective alternative to traditional TLC SSDs for GPU‑intensive workloads such as AI training.

The high‑bandwidth flash (HBF) concept pushes performance even further by placing flash dies directly on the same interposer as GPU high‑bandwidth memory. Up to 32 dies can be stacked using through‑silicon vias, creating a compact, low‑latency memory tier that excels at reads while keeping power draw modest. Although TSV manufacturing adds complexity, flash remains cheaper than DRAM, suggesting HBF could undercut HBM prices for inference and edge‑server applications. Kioxia is still evaluating TLC versus XL‑Flash for the stack, balancing cost against read speed.

Looking beyond the rack, Kioxia is prototyping an optical SSD that converts PCIe signals to light for transmission over distances up to 100 metres. This approach addresses the limitations of copper and Ethernet fabrics as PCIe 6 and future PCIe 7 standards demand terabit‑per‑second throughput. Backed by Japanese government funding and a network of silicon partners, the optical drive targets large GPU farms where rack‑to‑rack latency is critical. Combined with an extended Kioxia‑SanDisk joint venture and Western Digital’s shift to HDDs, the roadmap positions Kioxia as a key supplier for next‑generation AI infrastructure.

Kioxia eyes GPU memory stack with high-bandwidth flash push

Comments

Want to join the conversation?

Loading comments...