KIOXIA has announced its fastest SSDs to date with the GP1 series. According to the manufacturer, the PCIe 6.0 drives deliver up to 10 million random read IOPS and are designed to be accessed directly by GPUs. This is made possible by KIOXIA’s second-generation XL-FLASH flash memory. The series was unveiled on August 4, 2026, and the first evaluation samples are expected to be shipped to select customers by the end of the year.
10 Million IOPS: What KIOXIA’s GP1 Series Delivers
At first glance, this number may seem abstract, so here’s a quick context: Typical enterprise SSDs for data centers range from two to three million IOPS for random read operations. The GP1 series delivers three to four times that. This is measured at a block size of 512 bytes, which is critical for the intended use case. The drives are connected via PCIe 6.0 and NVMe 2.2.
It’s also interesting to note what KIOXIA is promising for the future: The architecture is expected to scale up to 100 million IOPS. This is explicitly a goal for future generations and not a specification for the first GP1. Anyone skimming the news might easily confuse the two figures.
It has been apparent for some time that PCIe 6.0 is slowly gaining traction in data centers. Just a few weeks ago, the Micron 9650 —the very first PCIe 6.0 SSD—entered mass production.
Second-Generation XL-FLASH and the 512-Byte Issue
Technically, the GP1 series differs significantly from traditional data center drives. Instead of TLC memory, it uses KIOXIA XL-FLASH, an SLC-based flash that the manufacturer positions as Storage Class Memory. Read latency is in the range of a few microseconds, whereas conventional NAND SSDs typically range from 40 to 100 microseconds.
The second point is access granularity. Standard SSDs operate with significantly larger blocks, whereas the GP1 series reads in 512-byte chunks. For AI workloads, this is exactly the right size: When offloading KV caches, performing embedding lookups, or traversing vector indexes, there are massive numbers of small, scattered read operations. An SSD optimized for sequential bandwidth wastes a large portion of its performance in these scenarios.
Incidentally, according to KIOXIA, power consumption per I/O also decreases compared to TLC-based SSDs. In racks where every watt is balanced against the GPUs, this is a significant consideration.
Flash as an Additional Storage Tier Alongside HBM
The real reason behind the GP1 series is the so-called “memory wall.” AI models are growing faster than High Bandwidth Memory (HBM) can be meaningfully scaled, because HBM is expensive, scarce, and thermally demanding. KIOXIA is therefore proposing an additional tier: fast flash directly below the HBM, which the GPU can access without going through the CPU.
This is exactly what KIOXIA has been working on since the spring. In March 2026, the company unveiled the GP series, developed in collaboration with NVIDIA as part of its Storage Next initiative and offering capacities of up to 25.6 TB. The GP1 series builds on this technology. Concepts such as Big Accelerator Memory (BaM) and NVIDIA’s BlueField-4 are moving in the same direction: the GPU retrieves data directly from the drive itself, rather than waiting for the host.
“The increasing performance and storage requirements of AI infrastructures call for a more comprehensive storage hierarchy that can efficiently and continuously supply GPUs with data,” explains Paul Rowan, Vice President and Chief Marketing Officer at KIOXIA Europe GmbH. The GP1 series reduces latency and CPU load while simultaneously increasing data throughput for AI, HPC, and data analytics workloads.
The reasoning behind this is simple: Flash is orders of magnitude cheaper per gigabyte than HBM. If this allows part of the data volume to be offloaded without the GPU sitting idle, it improves the utilization of expensive accelerators.
Form Factors, Cooling, and Endurance
The GP1 series is available in the EDSFF formats E3.S and E1.S with heights of 9.5 and 15 mm. For the 9.5-mm E3.S and E1.S models, KIOXIA supports a cold plate for liquid cooling, but all variants also run in traditional air-cooled servers. KIOXIA has already taken this approach with the NX1 series, the manufacturer’s first SSD with direct liquid cooling.
Noteworthy is the specified endurance of up to 50 DWPD. Read-intensive enterprise SSDs typically have an endurance rating of 1 DWPD, while mixed-use models have a rating of 3 DWPD. The GP1 series can therefore be overwritten fifty times its capacity daily—which is necessary for a cache layer that is constantly being refilled.
Technical Specifications of the KIOXIA GP1 Series at a Glance
| Feature | KIOXIA GP1 Series |
|---|---|
| Interface | PCIe 6.0, NVMe 2.2 |
| Storage | 2nd-generation KIOXIA XL-FLASH (SLC, Storage Class Memory) |
| Random Read | Up to 10 million IOPS at a 512-byte block size |
| Access granularity | 512 bytes |
| Form factors | E3.S, E1.S 9.5 mm, E1.S 15 mm |
| Cooling | Cold plate for E3.S and E1.S 9.5 mm; air cooling for all variants |
| Endurance | Up to 50 DWPD |
| Capacities | Not yet specified |
| Outlook | Architecture scalable up to 100 million IOPS, according to KIOXIA |
| Availability | Evaluation samples for select customers through the end of 2026 |
Availability, Unanswered Questions, and the Competition
Evaluation samples are scheduled to be distributed to select customers through the end of 2026. KIOXIA has not yet disclosed details regarding capacity, specific latency values, throughput, power consumption, or pricing. The GP1 series will be on display at the FMS trade show : The Future of Memory and Storage, taking place August 4–6 in Santa Clara. The series also received a Best-of-Show Award there in the Specialized Storage category.
KIOXIA is not alone in this field. SanDisk and SK hynix are pushing stacked NAND directly onto the accelerator with High Bandwidth Flash (HBF), while Samsung is promoting zHBM for HBM stacks above the chip. Four manufacturers, three mutually incompatible answers to the same question. Which approach prevails will likely be determined less by the spec sheets and more by what NVIDIA integrates into its platforms.
Just a few days earlier, KIOXIA had unveiled the CM10 series, the first enterprise SSDs with PCIe 6.0. The two series address different challenges: The CM10 uses BiCS FLASH to deliver capacities of up to 61.44 TB, while the GP1 focuses on access speed with a more modest capacity.
Conclusion on the KIOXIA GP1 Series
The KIOXIA GP1 series is not an SSD in the traditional sense, but rather an attempt to establish flash as a storage tier between HBM and DRAM. 10 million IOPS at 512-byte granularity are the key metrics here, even if plenty of questions remain regarding capacity, latency specifications, and pricing. By the end of 2026, we’ll see whether this approach pays off in real AI systems.
Sources: KIOXIA Europe, KIOXIA (press release)
