Xiaomi's new Xuanjie O3 chip just crushed the mobile SoC

Apple might be getting ready to drop the iPhone 18 Pro and its first foldable, but Xiaomi just decided to rewrite the entire script. At the recent Xuanjie technical briefing, Xiaomi unveiled its second-generation in-house SoC, the Xuanjie O3, and the numbers are frankly staggering. We are looking at a Geekbench 6.5 multi-core score of 15,000—that is nearly 40% higher than the current industry benchmark, Apple's A19 Pro. In extreme low-temperature AnTuTu testing, it even broke the 5-million mark, hitting 5.22 million.

However, the raw clock speed isn't the real story here. The architecture is where things get interesting for anyone following silicon design.

The Power of Ten Full Big Cores

To hit these theoretical peaks, Xiaomi opted for a massive 10-core CPU configuration: 6 "super" cores paired with 4 "big" cores, reaching a peak frequency of 4.35GHz. Instead of just chasing single-core clock speeds, the O3 is clearly betting heavily on multi-core throughput.

Scaling up to 10 high-performance cores in a mobile footprint is a massive thermal and spatial challenge. Using a 3nm process, Xiaomi increased the die size from 109mm² (on the O1) to 133mm². This allowed them to bump the transistor count from 19 billion to 24 billion.

But more cores usually mean more heat. To prevent this from becoming a battery killer, Xiaomi claims they've optimized the O3 to reduce power consumption by 25% across 80% of typical low-to-medium load scenarios. On the graphics side, the 16-core G2-Ultra NX GPU offers an 85% performance jump over the previous gen, with a massive 182% increase in ray tracing capabilities.

Solving the Memory Wall

If you have a massive CPU and GPU, you eventually hit the "memory wall"—the point where the processor is faster than the data can move. It’s like having a world-class chef in a kitchen with a tiny hallway to the pantry; the chef spends more time waiting than cooking.

Xiaomi tackled this with a few aggressive moves:

• LPDDR6 Support: They are pioneering 10667Mbps LPDDR6, pushing bandwidth to 113.8GB/s (nearly 50% higher than LPDDR5X).
• Massive Cache: Total on-chip cache has been expanded to roughly 60MB (including 12MB L2, 16MB L3, and 16MB SLC).
• Low Latency: Static memory access latency is down to 82ns, significantly beating the A19 Pro’s 92ns.

This is critical for on-device LLM agents. Every time a model generates a token, it has to pull billions of parameters from memory. To help with this, the O3 uses hardware-level lossless Huffman compression, which reduces memory bandwidth usage by 30% for Xiaomi’s MiMo quantized models.

Beyond Phones: The Xuanjie Ecosystem

The most shocking part of the announcement wasn't the mobile chip, but the specialized silicon designed for heavy AI workloads.

The Xuanjie O100 (AI Accelerator)
Unlike the general-purpose O3, the O100 is built purely for LLM inference. It uses a radical "Wafer on Wafer" 3D stacking technique with Hybrid Bonding. By vertically stacking NPU wafers and DRAM wafers, they've achieved a vertical interconnect with 2.58 million bonding nodes. The result? A memory bandwidth of 1.22TB/s—which is 16 times faster than the LPDDR5X found in flagship phones. It can hit local inference speeds of up to 330 tokens/s.

The Xuanjie D100 (Automotive AI)
For smart driving, the D100 uses a 3nm process and supports up to 160GB of unified memory. This is a huge deal for autonomous driving because it allows massive 200B-parameter models to run locally on the car, ensuring low latency even in tunnels or underground garages where cloud connectivity fails.

Xiaomi is essentially building a full-stack AI hardware vertical: HyperOS for scheduling, MiMo for the models, and Xuanjie for the silicon. They also teased the "Xiaomi AI Cube Prototype," a desktop AI workstation, and confirmed that the upcoming Xiaomi 18 Fold (rumored to have a wider "wide-fold" aspect ratio) will be powered by the O3.