Xiaomi unveiled three new Xuanjie chips yesterday, marking a significant expansion of its semiconductor ambitions beyond the smartphone sector. The triple launch positions the company's chip lineup as the foundational AI computing layer for its entire "human-vehicle-home" ecosystem, a strategic move that aligns with the rapid evolution of AI technology across multiple application scenarios.

Xuanjie O3: The Flagship AI SoC

The Xuanjie O3 is Xiaomi's second-generation high-end flagship SoC, maintaining the 3nm process node while boosting transistor count from 19 billion to 24 billion compared to its predecessor. According to Xiaomi's lab data, the chip achieves an impressive AnTuTu benchmark score of 5,228,014, making it the first mobile platform to break the 5 million point barrier in the industry.

On the CPU front, the Xuanjie O3 adopts a 10-core all-big-core architecture, featuring 6 super-large cores and 4 large cores with clock speeds up to 4.35GHz. This configuration delivers industry-leading baseline performance, with GeekBench 6 multi-core scores reaching 15,221, a substantial 60% improvement over the previous generation. The chip also maintains excellent energy efficiency in low-to-medium load scenarios, reducing daily power consumption by an additional 25%.

The graphics capabilities of the Xuanjie O3 have undergone a massive upgrade as well. It debuts the new G2-Ultra NX GPU with a 16-core ultra-large configuration, delivering an 85% improvement in traditional graphics performance over the O1, while ray tracing performance has surged by 182%. The GPU's energy efficiency has also been optimized, reducing power consumption by up to 64% at equivalent performance levels across all usage scenarios.

Beyond its powerful computing units, Xiaomi's Xuanjie team has focused on enhancing memory and storage configurations to accelerate data throughput and improve underlying operational efficiency. The Xuanjie O3 features a generous 60MB of on-chip cache across major modules, including a substantial 16MB SLC that addresses a previous generation shortcoming. It also debuts support for LPDDR6 memory with bandwidth reaching 113.8 GB/s, a 48% improvement. Xiaomi has innovatively applied a unified fusion bus architecture, achieving memory access latency as low as 82ns, setting a new industry benchmark.

The multimedia subsystem has also received comprehensive upgrades. The fifth-generation ISP supports single-camera resolutions up to 432 megapixels with 24-bit internal image signal processing, while AI RAW domain noise reduction for night video has been elevated to 4K/60FPS. The DPU extends screen zone tone mapping to more scenarios, enhancing dark-light readability and HDR content display. The VPU adopts a separate encoding/decoding architecture that accelerates video editing exports by 20% and introduces H.266 hardware decoding for the first time on Xiaomi phones.

The security module has been completely redesigned, earning both national cryptographic and CCRC EAL5+ certifications. In terms of baseband, communication performance is on par with mainstream flagship integrated solutions, while 5G scenario power consumption has been reduced by over 20% compared to the previous generation.

The Xuanjie O3 is not merely a flagship mobile SoC; it's an AI SoC fully embracing the AI era. AI hardware units are deployed across major internal cores to handle lightweight AI algorithms across various scenarios, minimizing NPU invocations and reducing cross-module communication latency and extra power draw.

The NPU has undergone deep hardware-level optimizations for large language models, with a 4-core architecture delivering 200 TOPS of computing power, far exceeding mainstream flagship offerings. The Xuanjie O3 also introduces the mobile industry's first SIMT-architecture Vector unit, providing industry-leading 3.13 TFLOPS of vector computing to address on-device large model hardware bottlenecks. Additionally, the NPU features expanded near-memory investment, combined with the 5-value quantization co-developed with Xiaomi MiMo and hardware-based Huffman lossless compression, enabling model inference speeds 45% faster than mainstream flagship SoCs.

Xuanjie O100: High-Bandwidth AI Chip for the Agent Era

The Xuanjie O100 is Xiaomi's self-developed 6nm on-device AI chip. Paired with the Xuanjie O3, it enables ultra-fast on-device large model inference, allowing devices to respond quickly to AI requests even in weak or no-network environments.

This chip employs the industry's most advanced 3D Wafer On Wafer stacking technology, using Hybrid Bonding to fuse two DRAM wafers and one NPU computing wafer at high temperatures. The Xuanjie O100 contains 2.58 million bonding nodes with a pitch of just 1.4μm, a specification that even surpasses the HBM memory used in cloud AI chips.

The ultra-high-density interconnection between computing and storage units delivers memory bandwidth 16 times that of mainstream phones, enabling on-device large model inference speeds of up to 330 tokens per second. This thoroughly resolves the bandwidth bottleneck that has constrained on-device large model computing.

The Xuanjie O100 also incorporates 14 NPU cores dedicated to large model processing, paired with Xiaomi's innovative XRING HB-Matrix high-bandwidth matrix bus, allowing efficient collaboration between internal cores for rapid AI task completion.

Xiaomi also addressed numerous manufacturing challenges through design innovation, including a six-month iteration on chip layout to overcome wafer warping and cracking issues during high-temperature bonding. The three-layer chip stack achieves face-to-face metal layer direct connection, with multiple patents already secured.

Xuanjie D100: High-Compute AI Chip for Intelligent Driving

The Xuanjie D100 is Xiaomi's self-developed 3nm intelligent driving chip with high AI compute power, also marking the industry's first 3nm autonomous driving chip. It contains 20 high-performance CPU cores and a powerful 16-core high-compute NPU.

Beyond efficiently running intelligent driving algorithms in vehicles, the Xuanjie D100 provides exceptional on-premises AI computing for individual users. A single chip supports up to 160GB of memory, enabling local deployment of large models exceeding 200 billion parameters. This eliminates expensive cloud API calls and physically prevents privacy data leakage risks.

Furthermore, the Xuanjie D100 supports multi-chip fusion computing through the Xuanjie high-speed interconnect bus, providing even greater aggregate hardware compute power for handling complex AI tasks locally.

Among the three new chips, the Xuanjie O3 will debut this September with the Xiaomi 18 Fold, while the O100 and D100 have completed development and will enter commercial use next year.

This triple launch signifies that Xuanjie has become the AI computing foundation for Xiaomi's entire ecosystem, spanning from pocket to cockpit and from living room to factory. A single Xuanjie hardware solution now covers the full spectrum of human-vehicle-home scenarios, meeting diverse AI demands for high energy efficiency, high bandwidth, and high compute power. More importantly, it demonstrates Xiaomi's steady transition toward becoming a hardcore technology leader with increasingly formidable chip design capabilities.