Kioxia Unveils PCIe 6.0 SSDs Targeting 10 Million IOPS for AI Workloads

Kioxia Introduces Three Enterprise SSD Series for AI Infrastructure

Kioxia unveiled three new Kioxia SSD families designed specifically for AI workloads at the Future of Memory and Storage conference in Santa Clara, California. The GP1 Series, CM10 Series, and NX1 represent the company's push into next-generation PCIe 6.0 NVMe SSDs and advanced cooling solutions for high-performance AI servers.1

1

2

2

According to Scott Nelson, Executive Vice President and Chief Marketing Officer for KIOXIA America, Inc., "AI is rapidly evolving across every industry, and memory is at the center of that transformation."2

2

GP1 Series Delivers 10 Million Random Read IOPS for GPU Direct Access

The GP1 Series represents Kioxia's attempt to bridge the AI memory wall by acting as high-speed capacity extension for GPU memory. Using second-generation XL-Flash storage-class memory, the GP1 delivers approximately 10 million random read IOPS in its initial architecture.1

1

4

4

Future generations target scaling toward 100 million IOPS, though this represents a long-term goal rather than current performance.1

1

The drives support 512-byte access granularity and reach endurance ratings of 50 drive writes per day, placing them far above ordinary read-intensive enterprise SSDs.1

1

Neville Ichhaporia, senior vice president and general manager of the SSD business unit at KIOXIA America, explained: "The AI memory wall is a critical challenge for our industry to solve on the way to growing the scale and capability of AI deployment. KIOXIA's XL-FLASH is uniquely positioned to deliver the extremely high performance and low-latency data access needed to support GPUs as a memory extension tier."4

4

The GP1 Series will be available in E3.S and E1.S 9.5 mm/15 mm form factors, with support for cold-plate liquid cooling. Evaluation samples are expected by the end of 2026.4

4

CM10 Series Achieves 92% Performance Gain With BiCS FLASH Generation 10

The CM10 Series marks Kioxia's first PCIe 6.0 enterprise SSDs built for AI inference and Key Value cache workloads. These drives feature BiCS FLASH generation 10 TLC NAND with 332 stacked layers, achieving the highest bit density and smallest chip size for 1Tb products.5

5

Compared to the previous-generation CM9 Series PCIe SSD, the CM10 delivers 92% higher sequential read performance and approximately 85% higher random read performance.2

2

5

5

The CM10 Series supports NVIDIA CMX context memory storage, expanding GPU memory on NVIDIA Rubin for AI inference.5

5

Capacities range from 1.60 TB to 61.44 TB, with read-intensive (1 DWPD) and mixed-use (3 DWPD) options. The drives are available in E3.S, E1.S, and 2.5-inch form factors, with E3.S and E1.S 9.5mm variants supporting direct cold-plate liquid cooling alongside traditional air cooling.5

5

The CM10 Series is NVMe 2.1 and OCP Datacenter NVMe SSD Specification 2.7 compliant.5

5

NX1 Series Brings Direct Liquid Cooling to PCIe 5.0 Storage

The NX1 represents Kioxia's first SSD line officially designed for cold-plate direct liquid cooling, addressing thermal challenges in GPU-equipped AI servers.3

3

This E1.S PCIe 5.0 NVMe SSD family succeeds the XD8 and introduces a new Kioxia controller alongside eighth-generation BiCS Flash TLC NAND using CMOS directly Bonded to Array architecture.3

3

Kioxia claims sequential-write performance is up to 38% higher than the previous XD8, with random writes improving by as much as 20%.3

3

The 9.5mm E1.S version operates under conventional airflow or connects to cold plates, while the thicker 15mm variant relies on air cooling. Capacities for one-drive-write-per-day models range from 1.92TB to 15.36TB.3

3

The NX1 supports the OCP Datacenter NVMe SSD Specification 2.6 and Flexible Data Placement, which lets software organize data to reduce internal write amplification and garbage collection.3

3

Direct liquid cooling becomes relevant as GPUs, CPUs and networking components increasingly consume most server airflow and thermal capacity. Extending the liquid loop to storage enables denser systems while reducing dependence on high-speed fans.3

3

Samples are already shipping to select hyperscale customers.3

3

What This Means for AI Infrastructure Development

These three product lines address different layers of storage for the AI Era. The GP1 Series tackles the High Bandwidth Memory cost barrier by providing GPU direct access at significantly lower cost per gigabyte than expanding HBM or DRAM capacity.4

4

The CM10 Series focuses on AI inference workloads where sequential and random read performance directly impact response times. The NX1 addresses thermal constraints in dense AI infrastructure.

Kioxia's approach reflects broader industry recognition that storage must evolve beyond traditional block storage paradigms. The GP1's 512-byte access granularity and extremely high random read IOPS position it as a memory tier rather than conventional storage.1

1

Watch for how software stacks adapt to leverage these capabilities, particularly in placing colder data on flash while reserving expensive memory for actively processed information. The technologies developed for enterprise AI infrastructure may eventually influence workstation and high-performance consumer SSDs.3

3