Overview of AI Chip Applications and Representative Companies

Based on Tsinghua University's "Artificial Intelligence Chip Research Report," this article offers a thorough introduction to AI chips, outlining their development status, mainstream enterprises, and usage scenarios.

1. AI Chip Application Fields

(1) Smartphones – Huawei’s Kirin 970 with Cambricon NPU (the world’s first mobile AI chip) and Apple’s A11 Bionic with a Neural Engine demonstrate on‑device deep learning for photography and image processing.

(2) ADAS (Advanced Driver‑Assistance Systems) – AI chips process massive sensor data (LiDAR, radar, cameras) enabling neural‑network‑based vehicle control and perception.

(3) Computer Vision Devices – Smart cameras, drones, facial‑recognition robots, and digital pens rely on local inference to avoid network latency.

(4) VR Devices – Microsoft’s HPU for HoloLens handles multi‑camera, depth, and motion sensor data, accelerating matrix and CNN operations for real‑time 3D reconstruction.

(5) Voice Interaction Devices – Companies like Qiying Tailun and Yunzhisheng provide chips with dedicated DNN accelerators for offline speech recognition and semantic understanding.

(6) Robots – Horizon Robotics and other firms deliver AI‑chip solutions for household and service robots.

2. Representative Domestic and International AI‑Chip Companies

Cambricon (中科寒武纪) – Founded in 2016, the first AI‑chip unicorn, offering both edge AI processors and cloud AI chips; notable product Cambricon‑1A.

UNISOC (中星微) – Established in 1999, introduced the world’s first NPU‑integrated SVAC video‑codec SoC for smart surveillance.

Horizon Robotics – Founded in 2015, develops BPU (Brain Processing Unit) IP for autonomous driving and smart vision.

DeepInsight (深鉴科技) – Founded by Tsinghua and Stanford researchers; its FPGA‑based DPU (Aristotle and Cartesian architectures) achieves up to 189× speedup over CPUs.

Tianjic (灵汐科技) – Develops next‑generation neural‑network processors supporting CNN, MLP, LSTM, and spiking neural networks, with toolchains for Caffe and TensorFlow.

Qiying Tailun (启英泰伦) – Provides ASIC‑based AI speech‑recognition chips (e.g., CI1006) optimized for DNN workloads.

Baidu – Released the 256‑core FPGA‑based XPU for accelerating its PaddlePaddle deep‑learning platform.

Huawei – Kirin 970 integrates a Cambricon NPU, delivering 1.92 TFLOPs FP16 performance with significant energy efficiency gains.

Nvidia – Pioneer of GPUs; its GPUs dominate AI training and inference workloads.

AMD – Offers CPUs, GPUs, and APUs, with Radeon Instinct accelerators for AI.

Google – Developed the TPU (Tensor Processing Unit) series, with TPU 3.0 reaching up to 100 PFlops.

Qualcomm – Integrates AI capabilities into Snapdragon SoCs and invests in AI‑focused startups.

Nervana Systems – Provides ASICs with high‑bandwidth memory for cloud AI services.

Movidius (Intel) – Supplies vision‑processing units (e.g., Myriad 2) for drones, AR/VR, and embedded vision.

IBM – Develops neuromorphic chips such as TrueNorth for brain‑inspired computing.

ARM – Introduced the DynamIQ architecture, enabling AI‑optimized cores and software libraries.

CEVA – Offers DSP IPs like CEVA‑XM4 and XM6 that support deep‑learning acceleration.

MIT/Eyeriss – Research project delivering a 168‑core CNN accelerator with 10× GPU performance.

Apple – A11 Bionic integrates a Neural Engine capable of 600 GFLOPs for on‑device machine learning.

Samsung – Actively developing AI chips for future smartphones and investing in AI‑chip startups.

For further technical details, readers are encouraged to scan the QR code and follow the official account to access additional resources.