Kioxia announces new Super High IOPS SSD that helps accelerate AI workloads on Nvidia GPUs — 25.6TB drive provides more GPU-accessible memory for faster data access

Nvidia and Kioxia have been working on a new SSD design that will keep Nvidia's AI GPUs fed with no downtime during intensive AI workloads. Kioxia has announced the new GP series, which Kioxia classifies as a "Super High IOPS" SSD. The new drive will be available for customers to evaluate by the end of 2026.

The new drive is part of Kioxia and Nvidia's move to bring incredibly fast storage to AI GPUs. The GP series comes with Kioxia's XL-flash, which is designed to achieve over 10 million IOPS, a figure that is around three to four times greater than traditional datacenter SSDs. XL-Flash is made using specialized NAND flash that boasts a read latency range of just 3 to 5 microseconds. By comparison, traditional SSDs normally have a peak at 3 to 4 million IOPS and have read latencies in the 40 to 100-microsecond range.

These new drives are part of Nvidia's Storage-Next design that sees servers utilize these drives in conjunction with a direct link to the GPU itself, bypassing additional latency penalties that would occur when transferring data from the CPU. This technology is designed to offset the limitations of existing HBM memory, giving Nvidia's AI GPUs a secondary cache layer to store data and keep the GPU cores operating at 100% without any downtime. This addresses the growing problem of growing AI models that are scaling towards trillions of parameters and context windows that consist of millions of tokens.

Article continues below

Follow Tom's Hardware on Google News, or add us as a preferred source, to get our latest news, analysis, & reviews in your feeds.

TOPICS

Aaron Klotz is a contributing writer for Tom’s Hardware, covering news related to computer hardware such as CPUs, and graphics cards.

2 Comments Comment from the forums

Stomx

Tens of terabytes, high speed, SLC... That is the only kind of electronics which will have some value even when the whole system ended in the city dump (if intensive cache access will not kill it)
Reply
Jame5

XL-Flash is made using SLC NAND flash (the fastest flash type available) that boasts a read latency range of just 3 to 5 microseconds. By comparison, traditional SSDs normally have a peak at 3 to 4 million IOPS and have read latencies in the 40 to 100-microsecond range.

Traditional SSDs started on SLC flash, and only transitioned to MLC, TLC, and QLC in recent years due to the customers showing the willingness to tradeoff latency and wear performance for raw capacity. The transition was made back before 3D stacking became a viable technology, and the industry at the time saw the only way to keep growing capacity was to move to more complex solutions.

But also: This raises interesting questions, like why 1-2TB SLC drives for consumers aren't available (at the same price as 4-8TB TLC/QLC drives). Some of us actually want the low latency. I still run my OS on an old Optane drive and use a PCIe 5.0 SSD for bulk secondary storage because of latency.
Reply