Add GIZMOCHINA as Preferred Source on Google
Xiaomi has taken another step in its self-developed chip plans, introducing three new Xring chips, one of which is a specialized accelerator meant to handle more powerful AI processing right on the device. Founder, chairman and CEO Lei Jun announced it at the company’s Autumn flagship product launch event.
The new AI accelerator chip, called the Xring O100, delivers up to 1.22TB/s of memory bandwidth. Xiaomi is calling it the world’s first 6nm 3D wafer-level stacked advanced packaging solution.
It uses an advanced hybrid bonding process, with the bonding pitch getting very close to the physical limit According to the company, this design gives it 16 times the memory bandwidth of traditional mobile phones and enables an on-device inference speed of 330 tokens per second.
The key idea is stacking the logic and high-speed DRAM dies vertically using wafer-on-wafer technology and hybrid bonding. That shortens the physical distance between the compute parts and the memory.
Xiaomi says this near-memory approach helps overcome the long-standing memory bottleneck that has limited how well large language models can run on edge devices. Even when regular mobile SoCs have decent compute power, they often get held back by slow data transfer. The O100’s high bandwidth and dense connections are meant to let on-device models generate tokens much faster.
It uses a multi-core NPU setup optimized for different stages of inference and supports dynamic data routing. Xiaomi has already shown prototypes, including tablet-style devices and “Xiaomi AI Cube” mini-PC form factors that pair the O100 with other Xring chips for sustained local AI workloads. Commercial use is expected in 2027, with potential applications in smartphones, tablets, robotics, and more.
As mentioned above, the O100 is part of Xiaomi’s broader Xring chip family, which also includes flagship mobile SoCs and chips aimed at cars. Instead of just chasing smaller process nodes, Xiaomi is focusing on bandwidth and packaging improvements. The aim is to make generative AI feel responsive, private, and usable even when you’re offline.
Don’t miss a thing! Join our Telegram community for instant updates and grab our free daily newsletter for the best tech stories!
For more daily updates, please visit our News Section.
(Source)
Comments