GPU-direct AI SSD

KIOXIA GP1: 10 Million IOPS PCIe 6.0 AI SSD

KIOXIA GP1 is an AI infrastructure SSD family built for GPU direct access and very high random-read performance. It is different from capacity flash: the point is not petabytes per drive, but a faster tier between GPU memory and general storage.

The 10 million IOPS figure must stay qualified. KIOXIA states up to 10 million random-read IOPS at a 512-byte block size, and says actual speed varies by host, software and workload.

Direct answer
GP1 is a PCIe 6.0 NVMe SSD using KIOXIA XL-FLASH generation 2 for GPU-direct AI storage and flash-based memory extension. KIOXIA says evaluation samples will be available to selected customers by the end of 2026. No public retail price is shown for GP1 today.
Share:WhatsAppFacebookPost on X
Current status
Evaluation samples
Selected-customer evaluation samples expected by end of 2026.
Interface
PCIe 6.0 / NVMe 2.2
As stated in KIOXIA GP1 announcement.
Flash
XL-FLASH Gen2
Low-latency flash for high random-read I/O.
Qualified performance claim
Up to 10M random-read IOPS
At 512-byte block size under KIOXIA's stated conditions.
GP1 availability and current price
SSDCapacityInterfaceNANDForm factorCommercial statusCurrent price
KIOXIA GP1
GPU-direct AI storage and flash-based memory extension
Capacity not publicly finalizedPCIe 6.0
NVMe 2.2
XL-FLASH generation 2E3.S / E1.SEvaluation
Evaluation samples are expected for selected customers.
Not yet available

What GP1 is

GP1 is not merely a faster version of a bulk enterprise SSD. It is a super-high-IOPS SSD architecture intended to let AI systems use flash as a lower-cost extension tier below GPU memory. That changes the conversation from $/TB alone to latency, IOPS, software path and GPU access model.

GPU HBM
Fastest and most expensive memory tier
โ†“
System memory
CPU-visible capacity for data preparation and service logic
โ†“
GP1 / XL-FLASH tier
Flash-based memory extension and GPU-direct access path
โ†“
Enterprise NVMe
General high-performance SSD storage
โ†“
Capacity storage
QLC flash, HDD or object-store tier for bulk data

GPU direct access

GPU direct access reduces CPU involvement in moving data between storage and the GPU path. For AI inference or analytics workloads that repeatedly fetch small blocks, the storage tier can become part of the performance-critical path rather than a passive storage device.

Why AI infrastructure might use flash as memory extension

HBM capacity is scarce and expensive. Very long context windows, KV cache growth, retrieval workloads and model-serving pipelines can benefit from a tier that is much larger than HBM and much faster for fine-grained access than ordinary bulk storage. GP1 belongs in that discussion, but it does not replace HBM.

GP1 vs CM10

Primary goalGP1 targets GPU-direct, super-high random-read IOPS. CM10 targets general enterprise and AI SSD performance.
Flash typeGP1 uses XL-FLASH generation 2. CM10 uses BiCS FLASH generation 10 TLC.
PCIe generationBoth include PCIe 6.0 positioning, but CM10's 2.5-inch configurations are listed as PCIe 5.0.
Capacity focusCM10 has published capacity tables up to 61.44TB. GP1's public story is IOPS and memory-extension architecture rather than finalized capacity.
AvailabilityCM10 is sampling to selected customers. GP1 evaluation samples are expected for selected customers by the end of 2026.

GP1 availability and price

KIOXIA describes GP1 as an evaluation-sample product for selected customers. No public public price is shown here today. If that changes, the price table will show a real offer rather than a manufactured MSRP.

Current enterprise NVMe alternatives

These are current enterprise SSD prices shown on DatacenterDisk. They are alternatives or reference points, not substitutes for unpriced announced products.

ProductCapacityInterfaceFormConditionPrice$/TBAvailability
SanDisk Ultrastar DC SN655 61.44TB U.3
WUS5EC0C1ESP7Y3
61.44TBNVMe-PCIe4U.3New$68,559.43$1115.88/TBIn stockBuy/View Price
Seagate Nytro 5060 E3.S 15.36TB
XP15360SE70045
15.36TBNVMe-PCIe4E3.SNew$5,879.00$382.75/TBIn stockBuy/View Price
Samsung PM9A3 7.68TB U.2 NVMe
MZQL27T6HBLA-00A07
7.68TBNVMe-PCIe4U.2Used$2,715.99$353.64/TBIn stockBuy/View Price
Samsung PM9A3 7.68TB U.2 NVMe
MZQL27T6HBLA
7.68TBNVMe-PCIe4U.2New$3,999.00$520.70/TBIn stockBuy/View Price
WD Ultrastar DC SN655 7.68TB U.3
WUS5EA176ESP7E1
7.68TBNVMe-PCIe4U.3New$8,690.03$1131.51/TBIn stockBuy/View Price
Seagate FireCuda X1070 4TB
ZP4000GS3A001
4TBNVMe-PCIe4M.2New$829.99$207.50/TBIn stockBuy/View Price
Micron 7400 PRO 3.84TB U.3
MTFDKCB3T8TDZ
3.84TBNVMe-PCIe4U.3Used$701.68$182.73/TBIn stockBuy/View Price
Micron 7450 PRO 3.84TB U.3
MTFDKCC3T8TFS
3.84TBNVMe-PCIe4U.3Used$1,990.00$518.23/TBIn stockBuy/View Price
Kioxia CD8 Series 3.84TB NVMe U.2
KCD8XVUG3T84
3.84TBNVMe-PCIe4U.2New$2,461.20$640.94/TBIn stockBuy/View Price
Micron 7600 PRO 3.84TB U.2 PCIe5
MTFDLAL3T8THG
3.84TBNVMe-PCIe5U.2New$2,599.00$676.82/TBIn stockBuy/View Price
PCIe 6 enterprise SSDs โ†’
Gen6 status hub with CM10 and GP1.
KIOXIA CM10 โ†’
General enterprise PCIe 6.0 SSD family.
245TB SSDs โ†’
Capacity flash for AI data lakes and object stores.
Largest SSD โ†’
Announced, shipping and current-price leaderboards.
AI storage โ†’
Model-storage tiers and architecture context.
AI memory CAPEX calculator โ†’
Model memory and storage budget impact.

Frequently asked questions

What is KIOXIA GP1?

KIOXIA GP1 is a PCIe 6.0 NVMe SSD family optimized for GPU direct access and flash-based memory extension in AI infrastructure. It is not just a larger bulk-storage SSD.

Does GP1 deliver 10 million IOPS?

KIOXIA states up to 10 million random-read IOPS at a 512-byte block size, with performance varying by host, software and conditions. The page includes that context with the claim.

Can I buy KIOXIA GP1 today?

KIOXIA says evaluation samples will be available to selected customers by the end of 2026. No public retail offer is shown here today.

How is GP1 different from CM10?

CM10 is a general enterprise PCIe 6.0 SSD family for high-performance storage. GP1 is built around super-high random-read IOPS, XL-FLASH and GPU-direct AI architectures.

Sources checked: KIOXIA GP1 announcement, KIOXIA CM10 announcement.