Under the backdrop of the widespread adoption of large AI models, the competition among terminal manufacturers in hardware foundations is entering deeper waters. At the Xuanjie chip technology briefing, Xiaomi officially launched its first dedicated AI acceleration chip for large models - Xuanjie O100, marking a critical step forward in its self-research and development of core computing chips.

Breakthrough Process and Powerful Bandwidth

In terms of underlying hardware architecture and manufacturing process, Xuanjie O100 adopts an advanced 6nm process technology and introduces a 3D wafer-level stacking solution, achieving high integration within a limited chip size. Thanks to this advanced architecture, the chip's memory bandwidth reaches up to 1.22TB/s, providing a solid physical foundation for efficient throughput of massive parameters.

Innovative In-memory Computing and Ultra-fast Inference Performance

In terms of architecture design, Xuanjie O100 innovatively adopts an in-memory AI computing architecture and is equipped with a dedicated Xuanjie high-bandwidth matrix bus, which can perfectly support the demand for multi-core parallel computing. In actual performance, when running the Xiaomi MiMo 3B model, the chip's inference speed reaches 330 Tokens/s, demonstrating excellent real-time response and processing capabilities for large models.