Kioxia Unveils UFS 5.0 Flash Memory to Speed Up On-Device AI

Kioxia America has introduced a new line of UFS 5.0 embedded flash memory, aimed at giving AI-enabled mobile and edge devices the storage speed, efficiency, and capacity needed to run increasingly demanding AI workloads locally. The company plans to demonstrate the technology next week at the FMS: Future of Memory and Storage conference, and commercial samples are now available in 512 GB and 1 TB capacities, with mass production expected by the end of 2026.

The new devices are built on the latest JEDEC UFS 5.0 standard, the industry specification governing embedded flash storage used in phones and similar devices, and use the newer MIPI M-PHY v6.0 physical interface along with the UniPro v3.0 protocol, including a mode called HS-Gear6, to support theoretical data transfer speeds of up to 46.6 Gb/s per lane. In practice, using two lanes together, that translates to roughly 10.8 GB/s of combined effective read and write performance.

The push behind faster storage comes down to how AI is increasingly running directly on devices rather than in the cloud. Large language models, multimodal AI systems, advanced imaging, and real-time inference all require moving much larger amounts of data quickly between storage and the processor. Kioxia says that up until now, UFS read speeds have been a bottleneck limiting how fast increasingly sophisticated on-device AI models can...