One report from Nikkei has suggested that Kioxia and NVIDIA are partnering up to push SSD performance into uncharted territory, as both names are eyeing a crazy 100M random IOPS, slated for 2027.
With today’s cutting-edge SSDs landing at 3M random 4K IOPS via PCIe 5.0 with Micron’s upcoming commercially available PCIe 6.0 drive topping out at ~5.5M, it’s kind of astounding for Team Green and ex-Toshiba-owned firm to come up with that sort of number. That’s not all, NVIDIA wants to put a pair to work with a GPU, therefore giving AI systems a combined 200M IOPS boost in data-crunching power.
This also follows the nature of how AI workloads actually behave, with inferencing and training relying heavily on small, random reads of data, with latency being the most prioritized aspect instead of sequential reads and writes. While standard SSDs use 4K blocks, AI servers benefit more from 512b blocks, which better reflect real-world usage and reduce response times.
According to Tom’s Hardware, it seems that Kioxia might be utilizing its XL-Flash tech, which includes single-level cell NAND with high endurance and ultra-low latency, with each die handling around 109,000 random read IOPS, which means hitting 100M would require close to a thousand of them working together, hinting at possible scaling challenges. And don’t forget about things like channel bandwidth, controller limits, and firmware overhead; all of them also contribute to real-life performance changes.
On the other hand, there’s High Bandwidth Flash (HBF) that stacks NAND dies with a logic die using TSVs and microbumps to massively boost parallelism. It still uses NAND, but in a more radical, GPU-like fashion. Whether Kioxia takes that route or sticks with XL-Flash, it’s clear the company is experimenting with next-gen flash technologies to break traditional limits.
But if they really did turn these goals into fruition, let alone by 2027, it’s going to spin out a different class of products, or maybe the “true industrial AI SSD for data centers”, and surely it’ll cost a hefty buck to acquire one, but money is always not going to be an issue if the tech world wants to keep the push strong.
