Xiaomi Unveils In-House SoC "Xuanjie O3" Alongside AI Acceleration Chip "O100"

Xiaomi Announces In-House SoC “Xuanjie O3”

Xiaomi announced its 2nd-generation in-house SoC, the “Xuanjie O3,” at the “Deep Tech Communication Conference” held on August 24, 2026. According to a report by Zheng Tingxu from iFengr, the chip recorded 15,000 points in the Geekbench 6.5 multi-core score, exceeding the current Apple A19 Pro by approximately 40%. Its AnTuTu extreme score reached 5.22 million points in a low-temperature lab environment.

The CPU of the Xuanjie O3 adopts a 10-core all-big-core design. It consists of six super big cores and four big cores, with a maximum clock speed reaching 4.35GHz. Manufactured on a 3nm process, the die area has expanded from the previous generation O1’s 109mm² to 133mm². The transistor count also increased from 19 billion to 24 billion, achieving a 26% improvement in density.

Memory Bandwidth and Latency Improvements

As chip performance improves, the issue of memory data transfer speed becoming a bottleneck becomes more apparent. The O3 supports the industry’s first 4x24-bit wide, 10,667Mbps LPDDR6 memory, directly boosting bandwidth to 113.8GB/s. This represents a performance improvement of about 50% compared to mainstream LPDDR5X.

The total internal cache of the chip has been expanded to approximately 60MB. CPU L2 accounts for 12MB, L3 for 16MB, and the system-level cache (SLC) for 16MB. Thanks to the restructured in-house integrated bus and physical upper-level wiring, the static access latency has been shortened to 82ns. This is lower than the Apple A19 Pro’s 92ns and the previous generation O1’s 121ns, respectively.

GPU Performance and Power Efficiency

The GPU features a 16-core “G2-Ultra NX.” According to official Xiaomi figures, graphics performance has improved by 85% compared to the previous generation, and ray tracing performance has surged by 182%. Eight NX neural network units are integrated within the GPU, providing 36TOPS of computing power.

What’s noteworthy is the improvement in power efficiency. At the same performance point, the O3’s power consumption is up to 64% lower than the O1. The CPU’s daily mid-to-low load coverage rate is 80%, and under these conditions, power consumption is reduced by 25% compared to the previous generation.

AI Inference Capability and Memory Optimization

The O3 is equipped with 200TOPS of A8W4 tensor computing power and 3.13TFLOPS of vector computing power. Dedicated hardware Huffman lossless compression technology for the Xiaomi MiMo 5-bit quantized model saves 30% of memory bandwidth usage. In conjunction with the 28MB of NPU-nearby memory, the first-response speed of MiMo 3B improved by 40%, and inference speed increased by 45%.

Xuanjie O100:

AI Inference Dedicated Acceleration Chip

Xiaomi further announced the “Xuanjie O100,” specialized for large model inference. It is manufactured on a 6nm process and adopts Wafer-on-Wafer (WoW) wafer-level 3D stacking technology. Two layers of high-speed AI dedicated DRAM wafers and one layer of high-performance NPU wafer are vertically connected, linked by 2.58 million physical bonding nodes.

This design allows the O100’s memory bandwidth to reach 1.22TB/s. This is equivalent to about 16 times that of the LPDDR5X used in current flagship smartphones. The 14 NPU cores utilize a Ring network and an All-to-All broadcast network flexibly, achieving local inference speeds of up to 330 Tokens/s.

Xuanjie D100: AI Chip for Autonomous Driving

For autonomous driving, Xiaomi has prepared the “Xuanjie D100.” Manufactured on a 3nm process, it features a 20-core CPU and a 16-core NPU. It supports up to 160GB of unified memory and can run models with a scale of 200 billion parameters locally in the vehicle. This enables real-time inference without relying on the cloud, even in scenarios with poor network connectivity like underground parking garages or tunnels. Commercialization is planned for 2027.

Editorial Opinion

In the short term, the debut of the Xuanjie O3 introduces a new competitive axis to the smartphone SoC market. Xiaomi achieving benchmark scores surpassing Apple and Qualcomm with its own design will further intensify the performance competition within the Android ecosystem. Over the next few months, other Android device manufacturers will likely be compelled to follow suit with similar all-big-core architectures and memory optimizations.

In the long run, the existence of the O100 and D100 is significant. Xiaomi deploying a cohesive chip strategy across three domains—smartphones, AI servers, and autonomous driving—demonstrates the deepening of the company’s vertical integration strategy. This approach of leveraging SoC design know-how cultivated in smartphones for AI acceleration and autonomous driving could influence computing architectures in the edge AI era.

An unverified point is the gap between benchmark scores and actual user experience. The 10-core all-big-core design theoretically boosts parallel performance, but the extent of improvement in practical terms for daily battery life and thermal management remains opaque. Furthermore, the 1…

References

• “小米发布玄戒 AI 芯片「全家桶」，还官宣了小米「阔折叠」”, by 郑廷旭 — 爱范儿, 2026-08-24T09:07:58.000Z (ARR)
• Source URL: https://www.ifanr.com/1676441?utm_source=rss&utm_medium=rss&utm_campaign=