Nvidia and Kioxia Collaborate on Revolutionary 100 Million IOPS SSD for AI Servers

Nvidia and Kioxia target 100 million IOPS SSD in 2027 -- 33 times more than existing drives for exclusive use in AI servers

Kioxia is working with Nvidia to build a solid-state drive that would deliver 100 million random IOPS already in 2027, the company said at a news conference earlier this month, Nikkei reports. Nvidia reportedly plans to use a couple of such SSDs -- totalling a whopping 200 million IOPS -- attached directly to its GPUs to boost AI performance. "We will proceed with development in accordance with the proposals and requests from Nvidia," said Koichi Fukuda, chief technology officer of Kioxia's SSD division. Kioxia's drives with 100 million random read speeds are projected to use a PCIe 7.0 interface to connect to GPUs in a peer-to-peer mode and will be exclusively designed for use in AI servers that need to access and process vast amounts of data quickly. Today's high-end solid-state drives can deliver around 3 million 4K random IOPS, but to meet the performance needs of modern and upcoming GPUs -- which are optimized for burst memory access -- they need to get substantially faster and change the way they interact with NAND media. Earlier this year, Silicon Motion's CEO Wallace Kou told Tom's Hardware that Nvidia was interested in building SSDs that offer as many as 100 million random IOPS, 33 times more than existing drives can deliver. Around the same time, Kioxia disclosed plans to build XL-Flash-based 'AI SSDs' with over 10 million 512K random read IOPS in the second half of 2026. AI workloads rely on frequent, small, random reads to retrieve embeddings, model parameters, or database entries. In such cases, 512-byte blocks better represent actual usage patterns than 4KB blocks and provide lower latency. While drives that operate 512B blocks may not deliver the same raw bandwidth as typical SSDs with 4K blocks, it is easier to scale out sequential read/write speeds with multiple drives than to lower the latencies of conventional SSDs. It remains to be seen whether Kioxia's AI SSD with a 10 million 512K IOPS random performance will materialize in 2026 if Kioxia plans to build drives with a 100 million IOPS random performance in 2027. What is interesting is how exactly Kioxia plans to build its 100 million IOPS drive. Its proposed AI SSD is based on XL-Flash, which is SLC NAND memory with high endurance, very low latency, and fairly high performance. Kioxia's XL-Flash devices feature 16 planes (up significantly from 3 to 6 planes on modern 3D NAND devices for client PCs), which points to higher sequential and random performance. As Kioxia does not publish specifications of XL-Flash, it is impossible to judge the per-device performance of this memory type. Meanwhile, considering that an Innogrit Tacoma-based 400GB XL-Flash SSD with 32 NAND dies (with seven allocated for overprovisioning) and a PCIe 5.0 x4 interface delivers 3.5 million random read IOPS and 0.5 million random write IOPS, we can estimate that each die contributes up to 109,375 random read IOPS and 15,625 random write IOPS -- though this calculation comes with some caveats. Assuming perfect linear scaling across loads of NAND devices, a 100 million 512B IOPS SSD would require 915 of such dies (presuming the same levels of over provisioning). Now that Kioxia knows how to pack 32 NAND ICs into a single package, it can certainly build a drive based on 915 XL-Flash ICs (in 28 packages). Such a drive would require a special controller with at least a PCIe 5.0 x16 host interface (though PCIe 7.0 x4 would be more preferable). The problem is, there is no perfect linear scaling. Real-world performance of NAND devices in SSDs is limited by channel bandwidth, multi‑plane constraints, command pipelining/overheads, queue depth, firmware, and loads of other factors. Hence, the best case scenario for a 100 million 512B IOPS SSD featuring XL-Flash devices is a multi-controller module with dozens of controllers and a switch. Such a solution may well make sense in all-flash arrays, but Kioxia is explicitly talking about an SSD. Since using traditional 3D NAND memory for a 100 million IOPS SSD with 512B blocks is not exactly feasible, whereas using a brand-new type of media on a commercial data center-grade product is highly unlikely, Kioxia might instead look to emerging technologies that use NAND memory in an unconventional way. One of such technologies is probably high bandwidth flash (HBF) that packs up to 16 NAND devices and a logic die (a controller?) into a single stack and interconnects them using TSVs and microbumps. While HBF layers still use proven NAND memory cells, they are organized in multiple arrays to achieve a very high level of parallelism and therefore performance. We do not know whether Kioxia plans to use HBF for the project or stick to something similar. However, it is safe to assume that the knowledge it will gain from its experimentation with HBF to build ultra-high-performance SSDs is something the company intends to leverage.

Kioxia-Nvidia project aims for SSD performance 33 times higher than today's top drives

Serving tech enthusiasts for over 25 years. TechSpot means tech analysis and advice you can trust. Forward-looking: The collaboration between Kioxia and Nvidia marks another step forward for AI data center infrastructure. By pushing toward SSDs capable of 100 million IOPS - dramatically higher than today's top-end drives, which peak around 3 million - the companies are tackling one of the biggest bottlenecks in training and deploying massive AI models: data movement. If successful, this breakthrough could not only accelerate the pace of innovation in GenAI, but also redefine how data centers are built, upend competitive dynamics in cloud and enterprise storage, and set new expectations for the entire hardware stack. Semiconductor memory maker Kioxia is developing next-generation SSD technology designed for ultra-fast read speeds to support demanding AI workloads. The company announced it intends to commercialize an SSD capable of delivering up to 100 million random IOPS by 2027, a performance benchmark about 30 - 35x greater than current high-end models. This ambitious project is being developed in collaboration with Nvidia. At a briefing in Tokyo, Kioxia explained that the new drive will connect directly to Nvidia GPUs, rather than routing through a server's central processor. This direct, peer-to-peer connection allows much faster data movement between storage and compute resources, a crucial advantage for large-scale AI models that depend on frequent, small, random data reads. Such tasks, including fetching embeddings and model parameters, place unique demands on memory and storage that are not handled efficiently by existing SSD configurations. Nvidia has also set an even more aggressive target: a two-SSD configuration delivering 200 million IOPS using the forthcoming PCI Express 7.0 standard, which supports high-speed, peer-to-peer GPU communication. For context, today's high-performance SSDs reach about 3 million IOPS on 4K blocks. Achieving a jump to 100 million will require major advancements in both NAND flash technology and interface architecture. One of the leading candidates to underpin Kioxia's SSD is its proprietary XL-Flash, a SLC NAND memory type engineered for high endurance, low latency, and strong performance. XL-Flash technology supports up to 16 planes within a NAND die compared to the 3 to 6 planes typical in consumer-grade 3D NAND. While Kioxia has not released full specifications, real-world data offers a glimpse of the challenge ahead. A 400GB XL-Flash SSD with 32 NAND dies and a PCIe 5.0 interface has demonstrated around 3.5 million random read IOPS in testing. If performance scaled perfectly (which it rarely does) a drive with 915 dies could theoretically hit the 100 million IOPS mark. In practice, however, scalability is limited by factors such as controller bandwidth, firmware overhead, and system architecture, meaning that multi-controller designs or modular SSD configurations may be necessary to reach the target. Recognizing the limits of traditional 3D NAND, Kioxia is also researching high-bandwidth flash (HBF), a new type of storage designed to combine the speed and scalability of high-bandwidth memory with far greater capacity. HBF combines up to 16 NAND chips and a logic die within a stacked module, interconnected using advanced packaging techniques to maximize parallelism and bandwidth. While it is unclear whether HBF will be part of the final product, this research signals Kioxia's broader strategy to deliver ultra-high-performance storage for the AI era.

NVIDIA rumored to team with KIOXIA to make new SSDs that are 100x faster for AI workloads

TL;DR: NVIDIA and KIOXIA plan to co-develop ultra-fast SSDs with nearly 100x faster read speeds by 2027, targeting AI servers to partially replace HBM as GPU memory expanders. This innovation leverages PCIe 7.0 and addresses growing AI-driven NAND demand, projected to reach 34% of the market by 2029. NVIDIA wants to co-develop new SSDs with KIOXIA that would be close to 100x faster in read speeds than current SSDs, and use them inside of AI servers to partially replace HBM as GPU memory expanders. In a new report from Nikkei, we're hearing that KIOXIA is looking to partner with NVIDIA to commercialize new SSDs by 2027 with nearly 100x faster read speeds, to use inside of AI servers and to partially replace HBM as GPU memory expanders. KIOXIA has previously said that by 2029, almost half of NAND memory demand is projected to be AI-related. Nikkei reports that NVIDIA is aiming for 200 million IOPS, with KIOXIA planning to use two SSDs to achieve that, which will also be on the next-next-generation PCIe 7.0 standard. Masashi Yokotsuka, Executive Vice President of KIOXIA said: "we will collaborate with the world's largest GPU manufacturer to achieve super performance in GPU systems". TrendForce reports news from a US investment firm report cited by TechNews that highlighted that the long-awaited comeback for NAND, which had missed the huge AI investment boom over the last two years. The report suggests surging AI inference demand and the need for high-speed, high-capacity storage with random I/O access are making QLC-based eSSDs the "go-to solution". Other emerging NAND products, including Nearline SSDs, could also gain traction as HDD supply tightens in late 2026/early 2027, while high-bandwidth flash (BHF) might help relieve some of the memory bottlenecks in AI clusters. Analysts from the company project AI-related NAND will make up 34% of the global NAND market by 2029, which will add around $29B in TAM.

NVIDIA Reportedly Partnering With Kioxia to Produce SSDs Up to 100x Faster Than Standard Models, Designed for AI Workloads

NVIDIA and its partners are pushing towards the next phase of 'AI memory,' and based on a new report, Kioxia plans to develop commercial SSDs that will be 100x faster than traditional ones. Well, it seems like NVIDIA is eying a replacement for HBM technology and is apparently looking towards improving conventional SSDs to the point where they can offer higher performance than modern-day HBM solutions. Now, in a report by Nikkei, it is revealed that Kioxia is partnering up with NVIDIA to develop AI SSDs that could be 100 times faster than conventional solutions, and it is intended to replace HBM by offering higher capacities, and by being mounted directly on the GPU. NVIDIA claims to want to achieve 200 million IOPS with the solution built by Kioxia, and the SSD manufacturer will likely utilize two different SSDs, each with 100 million IOPS, to achieve the objective. The solution is claimed to offer PCIe 7.0 connection to achieve the speeds NVIDIA desires, and more importantly, there is a need for a 'rework' from ground-zero to replace HBM with an SSD-like solution, and based on what Kioxia has disclosed in the past, we are likely looking at the entry of HBF (High-Bandwidth Flash) from the firm. HBF is the answer to the limitations posed by NAND memory, and SanDisk originally developed the standard. The key advantage of HBF is that it can offer immensely high capacities, scaling up to terabytes of memory capacity per device, which means that a traditional data center can leverage the large pool of memory for inferencing workloads. However, HBF isn't the only way Kioxia can achieve the objectives set by NVIDIA, rather the firm has its bet on solutions like XL-Flash, which is a is a high-performance NAND technology. It is interesting to witness the industry overcome the limitations present with HBM and scale up to newer solutions, and by the looks of it, NAND is going to play a vital role in the future of AI memory.