Nvidia, Kioxia Aim for 100M IOPS SSD by 2027 – 33x AI Server Boost
▼ Summary
– Kioxia is collaborating with Nvidia to develop a 100 million IOPS solid-state drive by 2027, which will connect directly to GPUs to enhance AI performance.
– These SSDs will use a PCIe 7.0 interface in a peer-to-peer mode and are designed exclusively for AI servers that require rapid data access and processing.
– The drive’s high performance targets 512-byte blocks to better match AI workload patterns, which rely on frequent small random reads for lower latency.
– Kioxia may utilize XL-Flash memory or emerging technologies like high bandwidth flash (HBF) to achieve the necessary parallelism and performance scaling.
– Challenges include the need for advanced controllers and overcoming limitations in NAND scaling, such as channel bandwidth and firmware constraints.
The race to power next-generation artificial intelligence has taken a significant leap forward with the announcement of a collaboration between Kioxia and Nvidia to develop a solid-state drive capable of delivering an unprecedented 100 million random IOPS by 2027. This ambitious project aims to directly attach these ultra-high-performance SSDs to GPUs, potentially offering a staggering 200 million IOPS per system and dramatically accelerating AI server capabilities.
Current high-end SSDs typically achieve around 3 million 4K random IOPS, making this new target a monumental 33-fold increase. Kioxia’s Chief Technology Officer for SSDs, Koichi Fukuda, confirmed the company is moving forward with development based on Nvidia’s specific proposals and requirements. These drives are expected to utilize a PCIe 7.0 interface and operate in a peer-to-peer connection mode with GPUs, tailoring them exclusively for AI workloads that demand rapid access to enormous datasets.
AI applications frequently rely on small, random read operations to fetch embeddings, model parameters, or database entries. In these scenarios, 512-byte blocks more accurately reflect real-world usage than standard 4KB blocks, offering substantially lower latency. Although drives optimized for smaller block sizes may not match the raw bandwidth of conventional SSDs, scaling sequential performance across multiple drives is often more feasible than reducing latency in existing designs.
Earlier this year, Silicon Motion’s CEO Wallace Kou revealed Nvidia’s interest in 100 million IOPS SSDs, while Kioxia separately announced plans for an “AI SSD” with over 10 million 512K random read IOPS by late 2026. It remains unclear whether both projects will proceed independently or if the 2026 offering will serve as a stepping stone toward the more ambitious 2027 goal.
A key question is how Kioxia intends to achieve such extreme performance levels. The company’s proposed AI SSD leverages XL-Flash, a type of SLC NAND memory known for high endurance, very low latency, and strong performance. XL-Flash devices feature 16 planes, a substantial increase over the 3 to 6 planes found in modern client-grade 3D NAND, suggesting significantly higher sequential and random throughput.
Extrapolating from existing XL-Flash drive performance, a 100 million IOPS SSD might require around 915 NAND dies, packaged into 28 modules. Such a design would demand a specialized controller with at least a PCIe 5.0 x16 host interface, though PCIe 7.0 x4 would be more suitable. However, real-world performance doesn’t scale perfectly due to limitations in channel bandwidth, multi-plane operation, command overhead, and firmware constraints. A more realistic architecture might involve a multi-controller module with dozens of controllers and an internal switch, a design more common in all-flash arrays than single SSDs.
Given the challenges of using conventional 3D NAND for such an extreme IOPS target with 512B blocks, Kioxia may turn to emerging technologies like high bandwidth flash (HBF). This approach stacks up to 16 NAND devices and a logic die into a single package, interconnected through TSVs and microbumps. While still using proven NAND memory cells, HBF organizes them into highly parallel arrays to achieve exceptional performance. Whether Kioxia adopts HBF or a similar technology, the knowledge gained from developing such ultra-high-performance SSDs will undoubtedly influence future memory innovations.
(Source: Tom’s Hardware)
