Xiaomi Introduces AI Computing Foundation with Three New Chips

Xiaomi has publicly introduced three artificial intelligence chips designed for device integration. The company presented these at its Xuanjie briefing. However, Xiaomi has not yet provided a definitive explanation for what constitutes an artificial intelligence phone.

Credit: UNSPLASH

The Xuanjie communication briefing recently detailed the O3, O100, and D100 chips. These components and their technical specifications together establish what Xiaomi describes as an "AI computing foundation" for its broader ecosystem, which includes human interaction, vehicles, and home environments. The Xuanjie O3 chip delivers 200 TOPS of NPU compute power. The O100 utilises 1.22 terabytes per second of bandwidth, intended to overcome what is termed the "memory wall" for large models processed directly on devices. And the D100 chip extends advanced computational capabilities to smart driving systems and local model inference operations.

This tripartite division of chips addresses distinct artificial intelligence requirements: high energy efficiency, substantial bandwidth, and significant compute capacity. On the software development front, the Xuanjie team collaborated with the MiMo model team. They developed a quantization scheme employing five values, which effectively compresses the average model bit width to 2.6 bits. This work resulted in a 45 per cent increase in decode performance on the MiMo 3 billion model, concurrently reducing power consumption by 26 per cent.

The O100 represents a particularly distinct technical development. This chip, built on a six nanometre manufacturing process, incorporates three dimensional stacking at the wafer level. This method vertically connects a DRAM wafer with an NPU compute wafer. This arrangement elevates bandwidth to 1.22 terabytes per second and facilitates up to 330 tokens per second during inference operations conducted directly on a device. The D100 chip is designated as a three nanometre chip for autonomous driving, capable of supporting up to 160 gigabytes of memory and local models with up to 200 billion parameters.

Commercial rollout plans indicate the O3 chip will ship later this year, accompanying the premium Xiaomi 18 Fold device. Both the O100 and D100 chips are scheduled for a commercial launch next year. Analysts observe that this event confirms Xiaomi's capability in designing silicon developed for artificial intelligence. But they also note the chips primarily demonstrate artificial intelligence operating at increased speeds and greater efficiency, rather than establishing these components alone as defining an artificial intelligence phone.

Xiaomi has refrained from offering a clear definition for such a device. A staff member informed Chinese media outlets that the company's "product foundation is sufficient". However, the staffer also indicated that designating a device as an artificial intelligence phone necessitates compliance and filing considerations. Xiaomi president Lu Weibing previously stated that an authentic artificial intelligence phone must introduce a substantial shift in human computer interaction, moving beyond merely faster chips.

Industry bodies have likewise grappled with this nomenclature. IDC, a market research organisation, defines a new generation of artificial intelligence phones as devices equipped with an NPU rated at 30 TOPS or more (INT8) for generative artificial intelligence processed on the device. The prevailing question concerns the role of these AI ready chips. A genuine artificial intelligence phone also requires agents operating at a system level, across various applications, and a meaningful alteration in user interaction. While a "memory wall" may have been surmounted, an "application wall" reportedly persists.

• Xiaomi unveiled three new artificial intelligence chips, the O3, O100, and D100 models, at its Xuanjie briefing.

• The O3 chip provides 200 TOPS of NPU compute, while the O100 achieves 1.22 terabytes per second of bandwidth through advanced stacking.

• The D100 is designed for autonomous driving, supporting extensive memory and large local models.

• Xiaomi has not provided a clear definition of an "AI phone," citing a need for significant human computer interaction shifts and regulatory compliance.

• Industry groups also face challenges in defining such devices, with IDC stipulating specific NPU performance requirements for new generation artificial intelligence phones. Source: Pandaily