Xiaomi Unveils AI Cube Prototype with 1.2TB/s Memory Bandwidth
Xiaomi has announced the "AI Cube" prototype, a three-chip mini-PC system designed for running large AI models locally. The system integrates the Xuanjie O3, Xuanjie O100, and Xuanjie D100 chips, delivering up to 1.22TB/s of memory bandwidth and supporting up to 160GB of RAM. This development signals a push towards high-performance edge computing, enabling consumers to run large language models (LLMs) locally without relying on cloud infrastructure. The massive memory bandwidth and RAM capacity address the primary hardware bottlenecks for local AI inference. The Xuanjie O100 chip utilizes an advanced 6nm process with 3D wafer-level vertical stacking (Wafer on Wafer) and Hybrid Bonding technology to achieve its high bandwidth. While the D100 chip (originally designed for electric vehicles) supports the 160GB RAM capacity, the exact architecture combining these memory pools remains unconfirmed.
## BACKGROUND
Large language models require significant memory bandwidth and capacity to run efficiently, as model parameters must be constantly loaded into memory during inference. Traditional consumer hardware often struggles with these workloads due to bandwidth limitations. Xiaomi's Xuanjie line represents their in-house silicon effort to optimize hardware specifically for these demanding AI tasks.