Xiaomi introduces xuanjie o100 AI chip with 6nm process and 3d stacking technology

Xiaomi has unveiled its first dedicated AI acceleration chip for large models, the Xuanjie O100, during a chip technology briefing. The launch marks a significant step in the company's self-research and development of core computing chips. The Xuanjie O100 employs an advanced 6nm manufacturing process and incorporates a 3D wafer-level stacking solution, enabling high integration within a compact chip size. This architecture allows the chip to achieve a memory bandwidth of up to 1.22TB/s, providing a robust foundation for handling large datasets efficiently. The chip also features an in-memory AI computing architecture and a dedicated high-bandwidth matrix bus, which supports multi-core parallel computing. In practical testing, the Xuanjie O100 demonstrated an inference speed of 330 Tokens/s when running the Xiaomi MiMo 3B model, showcasing its capability to deliver real-time performance for large AI models.