KIOXIA PCIe 6.0 SSDs Hit 10M IOPS for AI Storage

KIOXIA PCIe 6.0 SSDs Hit 10M IOPS for AI Storage

KIOXIA showcased a new generation of PCIe 6.0 NVMe SSDs and AI-oriented storage technologies at FMS 2026, highlighting how storage infrastructure is evolving alongside increasingly demanding AI workloads.

The headline product was the new KIOXIA GP1 Series, a PCIe 6.0 NVMe SSD capable of delivering up to 10 million random-read IOPS with 512-byte access sizes and optimized for near-GPU and GPU-direct workloads.

KIOXIA also demonstrated the CM10 Series PCIe 6.0 enterprise SSD, the liquid-cooling-capable NX1 Series, the latest BiCS FLASH Gen10 TLC NAND, high-capacity QLC storage, CXL memory modules based on XL-FLASH, and AI-focused software such as AiSAQ.

The demonstrations reflect a broader change in AI infrastructure design.

As AI systems move from model training toward large-scale inference and agentic workloads, storage increasingly needs to provide not only capacity but also high random-access performance, low latency, predictable throughput, power efficiency, and tight integration with accelerator-based compute.

🚀 KIOXIA GP1 Reaches 10 Million IOPS

#

The most notable product at the event was the KIOXIA GP1 Series, a PCIe 6.0 NVMe SSD designed for extremely high-performance AI workloads.

The GP1 can deliver up to 10 million random-read IOPS using 512-byte accesses.

This performance target is particularly relevant to workloads that repeatedly access relatively small pieces of data rather than processing large sequential files.

KIOXIA also optimized the GP1 for GPU direct access, positioning the SSD closer to the accelerator data path.

Why random IOPS matter for AI

#

AI storage performance cannot be evaluated solely through sequential bandwidth.

Large model checkpoints and datasets can generate substantial sequential traffic, but inference and retrieval-oriented workloads can introduce very different access patterns.

Examples include:

• Retrieval-augmented generation (RAG)
• Vector databases
• Feature stores
• Model serving
• KV-cache-related data movement
• Dataset indexing
• Small-object retrieval
• AI agent workloads

A simplified AI data path looks like:

Storage
   |
   v
NVMe SSD
   |
   v
Host CPU / Memory
   |
   v
GPU
   |
   v
AI Inference

Technologies such as GPU-direct storage aim to reduce unnecessary movement through the host CPU and system memory.

The goal is not simply to make the SSD faster in isolation, but to improve the efficiency of the complete storage-to-GPU data path.

⚡ PCIe 6.0 Moves Enterprise SSD Performance Forward

#

The GP1 and CM10 series demonstrate KIOXIA’s transition toward PCIe 6.0 storage for high-end enterprise and AI infrastructure.

PCIe 6.0 increases the available interface bandwidth substantially compared with previous PCIe generations, making it possible for storage controllers and NAND architectures to scale toward much higher throughput.

However, an SSD’s real-world performance depends on more than the PCIe interface.

The complete storage stack includes:

Application
    |
    v
Filesystem / Storage API
    |
    v
NVMe Driver
    |
    v
PCIe 6.0
    |
    v
SSD Controller
    |
    v
DRAM / Cache
    |
    v
NAND Flash

Performance can therefore be constrained by the controller, NAND media, firmware, queue depth, thermal conditions, workload characteristics, and host architecture.

This is particularly important for AI infrastructure, where storage devices can operate continuously under high I/O pressure.

🏢 KIOXIA CM10: Enterprise PCIe 6.0 Storage

#

KIOXIA also introduced the CM10 Series, a PCIe 6.0 enterprise NVMe SSD designed for modern data-center and AI workloads.

Compared with the previous-generation CM9 Series, KIOXIA reports:

The performance improvements target both high-throughput and high-IOPS workloads.

That combination is increasingly important in AI servers because the storage subsystem may simultaneously handle model loading, checkpoint operations, dataset access, logging, retrieval workloads, and other background traffic.

BiCS FLASH Gen10

#

The CM10 is also notable because it is the first SSD highlighted by KIOXIA to use its latest 10th-generation BiCS FLASH TLC technology.

The NAND features a 332-layer stacked structure.

KIOXIA is also using CMOS Direct Bonded Array (CBA) technology together with additional lateral scaling techniques.

The basic objective of increasing vertical layer count is straightforward: place more storage cells into a given physical area.

However, higher layer counts introduce additional manufacturing and integration challenges. Advanced NAND scaling therefore increasingly depends on a combination of vertical stacking, lateral shrink, CMOS integration, and packaging improvements rather than layer count alone.

💧 NX1 Brings Liquid Cooling to SSDs

#

Another notable demonstration was the KIOXIA NX1 Series, a data-center NVMe SSD designed to support direct liquid cooling.

Liquid cooling has traditionally been associated primarily with CPUs and GPUs.

As accelerator and storage performance increases, however, SSDs are also becoming important contributors to rack-level thermal density.

A simplified high-density AI server may look like:

+--------------------------------------+
|              AI Server               |
|                                      |
|  GPU  GPU  GPU  GPU                  |
|  GPU  GPU  GPU  GPU                  |
|                                      |
|  CPU       Memory       NIC           |
|                                      |
|  NVMe SSDs / Storage                 |
+--------------------------------------+
           |
           v
     Cooling Infrastructure

When multiple components operate at high utilization, thermal management becomes a system-level problem.

Direct liquid cooling for SSDs allows storage devices to participate in the same high-density thermal-management strategy as other major components.

This can become increasingly important as PCIe 6.0 SSDs push higher performance within constrained server chassis.

🧠 Storage Becomes Critical as AI Moves Toward Inference

#

KIOXIA’s keynote, “Reshaping Data Center Storage in the AI Era with Flash Memory and SSDs,” focused on a major shift in AI infrastructure.

The industry is gradually moving from an environment dominated by model training toward one where inference becomes a continuously operating production workload.

Training tends to generate enormous sequential data flows, including:

• Dataset ingestion
• Checkpoint writing
• Model checkpoint loading
• Distributed training traffic
• Intermediate data processing

Inference introduces another set of requirements.

Production AI systems may need to continuously retrieve:

• Model weights
• Context data
• Documents
• Embeddings
• Vector database records
• User-specific information
• Agent state
• Cached data

This changes the performance profile expected from storage.

From bandwidth to data-access efficiency

#

The traditional storage question was often:

How many gigabytes per second can the storage system deliver?

AI infrastructure increasingly requires additional questions:

• How quickly can small objects be retrieved?
• How low is tail latency?
• How efficiently can data reach GPUs?
• How much CPU overhead does storage generate?
• How much power does each I/O operation consume?
• How does performance scale across thousands of drives?

This is why KIOXIA’s portfolio spans both extremely high random IOPS and high-capacity flash storage.

🔗 Balancing Memory and Storage for AI Factories

#

KIOXIA also co-hosted an NVIDIA panel titled “Balancing Memory and Storage to Power Large-Scale AI Factories and AI Agents.”

The subject highlights an increasingly important issue in AI system architecture: memory and storage can no longer be designed independently.

A modern AI server may contain several levels of the memory hierarchy:

Fastest
   |
   v
GPU SRAM / Registers
   |
   v
GPU HBM
   |
   v
System DRAM
   |
   v
CXL Memory
   |
   v
NVMe SSD
   |
   v
High-Capacity Storage
   |
   v
Slowest / Largest

Each layer provides a different balance of latency, bandwidth, capacity, and cost.

As model sizes and context windows expand, workloads increasingly require intelligent movement of data across this hierarchy.

The objective is not to eliminate storage access entirely.

Instead, system architects need to ensure that frequently accessed information remains in fast memory while larger or less frequently accessed datasets reside on increasingly capable storage.

📦 245.76 TB QLC SSDs Target Large AI Datasets

#

KIOXIA also demonstrated its LC9 Series, including a 245.76 TB NVMe SSD based on QLC flash.

QLC stores four bits per NAND cell, enabling significantly higher density than TLC.

The primary advantage is capacity.

For AI infrastructure, high-capacity SSDs can reduce the number of drives required to store large datasets, model repositories, checkpoints, and other data.

A simplified capacity-density comparison is:

TLC
  |
  +--> Higher endurance / performance characteristics
  |
  +--> Lower density than QLC

QLC
  |
  +--> Higher storage density
  |
  +--> Lower cost per bit
  |
  +--> Attractive for capacity-oriented workloads

The trade-off means QLC is not intended to replace high-performance TLC across every workload.

Instead, different NAND types can occupy different positions in the AI storage hierarchy.

🔍 AI RAG and Vector Database Scaling

#

KIOXIA also demonstrated AiSAQ, its open-source software technology aimed at improving the scalability of vector databases for AI RAG workloads.

Retrieval-augmented generation requires an AI system to retrieve relevant information before generating a response.

A simplified RAG pipeline is:

User Query
    |
    v
Embedding Generation
    |
    v
Vector Search
    |
    v
Relevant Documents
    |
    v
LLM Context
    |
    v
Generated Response

As document collections grow, the vector database can become a major storage and retrieval bottleneck.

This makes storage performance increasingly relevant to inference latency.

The challenge becomes even more pronounced when AI applications need to search massive datasets while simultaneously serving many concurrent users or agents.

High-capacity flash combined with efficient retrieval software can therefore become an important part of the RAG infrastructure stack.

🧠 CXL Memory and XL-FLASH Expand the Memory Hierarchy

#

KIOXIA also demonstrated CXL memory modules using XL-FLASH.

Compute Express Link (CXL) provides a standardized interconnect for attaching memory resources to processors and accelerators.

This creates additional opportunities to expand the memory hierarchy beyond traditional local DRAM.

The resulting architecture can look like:

CPU / Accelerator
       |
       +---- Local DRAM
       |
       +---- CXL Memory
       |
       +---- NVMe SSD
       |
       +---- High-Capacity Flash

KIOXIA’s XL-FLASH technology targets the gap between conventional DRAM and higher-capacity NAND storage, providing another tier for system architects to consider.

For AI systems with large working sets, such intermediate memory tiers can potentially reduce pressure on both expensive high-bandwidth memory and slower storage.

📱 UFS 5.0 Extends AI Storage Beyond Data Centers

#

The FMS 2026 demonstrations were not limited to data-center hardware.

KIOXIA also showcased UFS 5.0, positioning the technology as a high-performance and AI-compatible flash-storage solution.

UFS is particularly relevant to mobile and embedded systems, where power efficiency, compact packaging, and storage performance must coexist within strict physical constraints.

As on-device AI becomes more capable, local storage increasingly needs to accommodate:

• AI models
• Embeddings
• User data
• Multimodal content
• Application assets
• Local inference caches

This creates a similar architectural trend across data centers and edge devices: AI increasingly requires local access to large quantities of data.

🧱 BiCS FLASH Gen10 and Advanced NAND Scaling

#

Beyond SSD products, KIOXIA highlighted its BiCS FLASH Gen10 technology.

The new generation combines:

• 332-layer 3D NAND
• TLC flash
• CMOS Direct Bonded Array technology
• Lateral scaling improvements
• Higher-density packaging

NAND scaling is increasingly a multi-dimensional engineering problem.

Vertical stacking increases the number of layers, while lateral scaling increases the number of cells that can fit within the same horizontal area.

CBA technology further separates and integrates CMOS and array structures to improve scaling flexibility.

Together, these techniques allow flash manufacturers to continue increasing density without relying on a single scaling dimension.

📦 32-Die BGA Packages Increase Flash Density

#

KIOXIA also demonstrated high-capacity, low-latency QLC flash using 32-die stacked BGA packages.

Advanced packaging is becoming increasingly important as NAND density increases.

Instead of improving capacity exclusively at the individual die level, manufacturers can also increase the number of dies integrated into a single package.

This creates a hierarchy of density improvements:

NAND Cell
   |
   v
NAND Die
   |
   v
Multi-Die Package
   |
   v
SSD
   |
   v
Storage Server
   |
   v
AI Data Center

Each level introduces different constraints involving power delivery, thermal management, signal integrity, controller architecture, and manufacturing complexity.

🏗️ KIOXIA’s FMS 2026 Demonstrations

#

KIOXIA’s two-story exhibition booth brought these technologies together through a series of demonstrations.

The key demonstrations included:

Rather than presenting these technologies as isolated products, the demonstrations collectively illustrate KIOXIA’s approach to the emerging AI storage stack.

⚙️ Storage Is Becoming Part of the AI Accelerator Architecture

#

The most important takeaway from KIOXIA’s FMS 2026 presence is that storage is increasingly being designed as part of the AI compute architecture, rather than treated as a peripheral subsystem.

The conventional model was relatively simple:

CPU
 |
 +--> DRAM
 |
 +--> Storage

Modern AI infrastructure is considerably more complex:

                    +--> HBM
                    |
GPU / Accelerator --+--> System Memory
                    |
                    +--> CXL Memory
                    |
                    +--> NVMe SSD
                    |
                    +--> High-Capacity Flash

As AI models become larger and inference becomes more data-intensive, the boundary between compute, memory, and storage continues to blur.

The highest-performing systems will therefore depend not only on faster GPUs but also on efficient movement of data across the entire hierarchy.

🔮 Conclusion: PCIe 6.0 SSDs Target the Next AI Bottleneck

#

KIOXIA’s FMS 2026 portfolio demonstrates how quickly enterprise storage is adapting to AI infrastructure requirements.

The GP1 Series pushes random-read performance to 10 million IOPS, targeting near-GPU and GPU-direct workloads. The CM10 Series brings PCIe 6.0 to enterprise SSDs while delivering substantial performance gains over the previous CM9 generation. The NX1 Series addresses thermal constraints with direct liquid cooling.

At the flash level, BiCS FLASH Gen10 introduces a 332-layer TLC architecture alongside CBA and additional scaling technologies, while high-capacity QLC products such as the 245.76 TB LC9 target increasingly large AI datasets.

Meanwhile, CXL memory, XL-FLASH, UFS 5.0, AiSAQ, and advanced multi-die packaging demonstrate that KIOXIA is approaching AI storage as a complete hierarchy rather than a single SSD product category.

The larger industry trend is clear.

As AI systems transition from model training toward continuous inference, RAG, agentic workloads, and large-scale production deployment, storage latency, IOPS, capacity, power efficiency, and data-path efficiency are becoming first-class AI infrastructure concerns.

The next generation of AI performance will therefore not be determined by accelerators alone.

It will increasingly depend on how efficiently the entire system can move data from flash to memory and ultimately to the compute engines that consume it.