Xiaomi Unveils Xuanjie O100 AI Chip with 6nm 3D Wafer-Level Stacking
Xiaomi unveiled the Xuanjie O100, an in-house AI acceleration chip utilizing the world's first 6nm 3D wafer-level stacked advanced packaging. The chip delivers a high memory bandwidth of 1.22 TB/s and achieves an edge AI inference speed of 330 tokens per second (TPS). This breakthrough significantly alleviates memory bandwidth bottlenecks for on-device AI, enabling fast model execution directly on edge hardware. It also marks a major milestone in Xiaomi's internal chip architecture and advanced semiconductor packaging capabilities. The Xuanjie O100 utilizes an advanced hybrid bonding process with interconnect pitch approaching physical limits, yielding 16 times the memory bandwidth of traditional smartphone memory. This massive throughput is the key driver enabling its 330 TPS edge AI inference speed.
## BACKGROUND
Advanced 3D wafer-level packaging and hybrid bonding enable semiconductor dies to be vertically stacked with dense interconnects, dramatically boosting bandwidth while minimizing physical footprint. On-device edge AI relies heavily on such high memory bandwidth to stream complex neural network weights into processing cores rapidly without latency from cloud servers.