Xring O100 chip hits 330TPS with 1.22TB/s memory bandwidth

Xiaomi’s Xring O100 AI accelerator chip is designed around memory speed as much as compute, and that is the key detail here. The 6nm part uses 3D wafer-level stacked packaging, and Xiaomi says that helps it reach 330 tokens per second in demo claims while pushing memory bandwidth to 1.22TB/s.

That matters because on-device AI often slows down when data cannot move fast enough between logic and memory. Xiaomi is now pointing to hybrid bonding and wafer-on-wafer stacking as the answer, while also showing how the Xring chip family could extend beyond phones into tablets, mini-PCs and robots.

Hybrid bonding process and wafer-on-wafer stacking cut the memory bottleneck

Xiaomi says the bonding pitch is close to the physical limit, which is the point of the hybrid bonding process here. The idea is straightforward even if the engineering is not: reduce the distance between compute and memory so the chip spends less time waiting on data.

For on-device AI, that kind of change can matter as much as raw compute. Large language models need a steady flow of data to stay responsive, and bottlenecks in memory transfer can make them feel slow even when the chip itself is capable. Xiaomi’s pitch is that the Xring O100 is built to avoid that slowdown.

The company’s demo claims also frame the chip as a direct challenge to the edge-AI limits of conventional smartphone SoCs. In practice, that means Xiaomi is not just talking about faster processing, but about a more balanced system where memory architecture helps keep inference moving.

Xring chip family for AI moves beyond phones to tablets, mini-PCs and robots

Lei Jun showed prototypes that pair the accelerator with other Xring chips in tablet-style devices and the Xiaomi AI Cube mini-PC. That makes the announcement bigger than a single chip benchmark, because it suggests Xiaomi is building a broader hardware family around local AI tasks.

The company’s direction appears to be offline generative AI across several device types, not only smartphones. Tablets, mini-PCs and robotics were all part of the demonstration story, which points to a more general push for sustained local inference speed.

Commercial use is expected in 2027, so this is not a near-term consumer launch. Still, the timeline matters because it gives Xiaomi room to turn the Xring concept into a wider platform rather than a one-off silicon experiment.

For now, the main takeaway is simple. Xiaomi is placing memory bandwidth at the center of its AI hardware strategy, and the Xring O100 is the clearest sign yet of that approach. The 330TPS claim is the attention-grabber, but the underlying story is the packaging and memory design that made it possible.