Nvidia and Kioxia Collaborate on Revolutionary 100 Million IOPS SSD for AI Servers

Nvidia and Kioxia's Ambitious SSD Project

Nvidia and Kioxia are joining forces to develop a groundbreaking solid-state drive (SSD) that promises to deliver an astounding 100 million random IOPS by 2027 1

1

. This collaboration aims to create storage solutions that are 33 times faster than existing high-end drives, which currently peak at around 3 million IOPS 2

2

.

Revolutionizing AI Server Infrastructure

The primary goal of this project is to address one of the most significant bottlenecks in training and deploying massive AI models: data movement. These ultra-fast SSDs are designed exclusively for use in AI servers, where they will be directly connected to Nvidia's GPUs using a peer-to-peer mode 1

1

. This direct connection allows for much faster data transfer between storage and compute resources, a crucial advantage for large-scale AI models that rely on frequent, small, random data reads 2

2

.

Technological Innovations

To achieve this unprecedented performance, Kioxia is exploring several cutting-edge technologies:

1. XL-Flash: A proprietary SLC NAND memory type engineered for high endurance, low latency, and strong performance. XL-Flash supports up to 16 planes within a NAND die, compared to the 3 to 6 planes typical in consumer-grade 3D NAND 2

   2

   .

2. High Bandwidth Flash (HBF): This emerging technology packs up to 16 NAND devices and a logic die into a single stack, interconnected using TSVs and microbumps. HBF layers use proven NAND memory cells organized in multiple arrays to achieve a very high level of parallelism and performance 1

   1

   .

PCIe 7.0 and Future Prospects

The new SSDs are projected to use the forthcoming PCIe 7.0 interface, which supports high-speed, peer-to-peer GPU communication 2

2

. Nvidia's ambitious target is to achieve 200 million IOPS using a two-SSD configuration 3

3

.

Impact on AI and Data Centers

This breakthrough could potentially redefine how data centers are built, upend competitive dynamics in cloud and enterprise storage, and set new expectations for the entire hardware stack 2

2

. The new SSDs might partially replace HBM as GPU memory expanders, addressing the growing AI-driven NAND demand, which is projected to reach 34% of the market by 2029 3

3

.

Challenges and Future Development

While the potential of these new SSDs is immense, there are significant challenges to overcome. Achieving perfect linear scaling across loads of NAND devices is not feasible, and real-world performance is limited by various factors such as channel bandwidth, multi-plane constraints, and firmware 1

1

. As the project progresses, Kioxia and Nvidia will need to address these challenges to bring their revolutionary SSD to market by 2027.