Staff Reporter |
Editor | Wen Shuqi
On August 24, Xiaomi held a technical briefing for its Xuanjie chips, unveiling the detailed specs and product roadmap for the Xuanjie O3 flagship SoC, the Xuanjie O100 AI accelerator, and the Xuanjie D100 smart-driving chip.
These three chips target consumer electronics main processors, AI inference acceleration, and automotive intelligent driving, respectively. This signals that Xiaomi’s self-developed chip initiative is now pushing forward with its platform-based strategy, rolling out across multiple product lines in parallel.
The first to face market scrutiny will be the Xuanjie O3. According to Zhu Dan, Xiaomi’s Vice President and head of chip business, the O3 is built on a 3nm process, with a die size of 133mm² and 24 billion transistors—a 26% increase over the previous generation.
On the CPU side, it features a 10-core architecture with 6 super-large cores and 4 large cores, clocked up to 4.35GHz. Xiaomi claims CPU performance is up 60% compared to the previous generation, while the 16-core GPU delivers an 85% boost.
Unlike traditional flagship SoC battles that revolve mainly around CPU and GPU horsepower, Xiaomi this time put a much stronger emphasis on the O3’s targeted optimization for on-device AI. Its NPU adopts a quad-core, low-power design, delivering 200 TOPS of Tensor compute and 3.13 TFLOPS of Vector compute.
“The NPU is our biggest upgrade this time—it’s an architecture deeply optimized for large language models,” Zhu Dan told media including NUPIAO.
Xiaomi also teamed up with its MiMo team to optimize the chip at the model level. In tests with a 3B-parameter MiMo model, the Xuanjie O3 showed a 40% improvement in Prefill performance and a 45% gain in Decode performance compared to traditional flagship SoCs.
In terms of product rollout, Xiaomi’s foldable flagship, the Xiaomi 18 Fold, will be the first to feature the Xuanjie O3, and the Xiaomi Pad 9 Pro Max will also be powered by the same chip.
Ivan Lam, senior analyst at Counterpoint, told NUPIAO that based on the specs disclosed so far, the Xuanjie O3 sits comfortably in the top tier of 3nm flagship SoCs. “Of course, to truly compare it with Qualcomm and MediaTek, we’d need to put it side by side with same-generation, same-process products.”
Xiaomi also unveiled the Xuanjie O100 AI accelerator at the briefing. Built on a 6nm process with 3D wafer-level stacking, it offers 1.22TB/s of bandwidth. Architecturally, the O100 adopts a near-memory AI compute design, integrating 14 dedicated NPU cores for large models, and leverages Xuanjie’s high-speed interconnect bus to boost data transfer efficiency.
Xiaomi says that when running the Xiaomi MiMo 3B model, the O100 can achieve inference speeds of up to 330 Tokens/s.
The company also introduced a 3nm automotive smart-driving chip, the Xuanjie D100, which packs a 20-core CPU and a 16-core NPU, developed for high-level intelligent driver-assistance scenarios. Additionally, the D100 can deliver massive on-device AI compute for individual users, supporting up to 160GB of memory per chip—enough to locally deploy large models with over 200 billion parameters.
With this, Xiaomi is attempting to build a comprehensive self-developed compute ecosystem spanning consumer electronics and smart vehicles. Beyond running a dual-track strategy of external procurement and in-house chip development, the real value of self-developed chips for a company juggling smartphones, IoT products, and automotive lines lies in the room it creates for deep software-hardware co-optimization.
Ivan Lam pointed out that in today’s AI-centric landscape, a chip’s AI capability has become the top priority—and AI is demanding increasingly complex things from silicon. All three chips Xiaomi announced today lean heavily into AI, which is also about expanding the application scenarios for its in-house silicon. From there, as we head into the AI terminal era, Xiaomi’s Xuanjie chip series, HyperOS, and MiMo self-developed models can all be jointly optimized around the same on-device AI capabilities, potentially securing a system-level edge spanning models, software, and hardware.
But whether this strategy can actually translate into commercial competitiveness remains to be seen in the market. Developing advanced-node chips demands long-term investment, and for AI accelerators and automotive-grade chips, the actual scale of deployment will ultimately determine their economic impact.
At this briefing, Xiaomi didn’t disclose figures that would better illustrate the market performance of the previous-generation Xuanjie O1. However, Zhu Dan noted that as the mainland’s first 3nm flagship SoC, the O1 has “firmly established itself in the top tier” in terms of performance.
