Artificial Intelligence IC "refers to an integrated circuit specifically designed for efficient operation of artificial intelligence algorithms.

Simply put, it is the 'brain' that enables various AI applications to run on hardware.

Currently, AI chips are developing along two main lines: cloud side and end side

1. Cloud side AI chip: focusing on ultimate performance

Mainly deployed in data centers, responsible for handling AI model training and large-scale inference tasks.

·GPU: With its powerful parallel computing capabilities, represented by NVIDIA H100 and B200, it is currently the de facto standard for AI training.

·ASIC: Specialized chips customized for specific scenarios, such as Google's TPU and various neural network processors (NPUs), which are more efficient in inference tasks.

·Frontier exploration: Using wafer level chips (such as Cerebras) to break area limitations, or using photons to replace electronic computing, to achieve extremely low latency optical chips.

2. End side AI chip: Pursuing low power consumption and security

Deploy on terminal devices such as mobile phones, headphones, and cars, emphasizing the completion of inference under low power consumption.

·AI MCU: A microcontroller integrated with an AI acceleration unit that can locally run sensor data processing or generative AI on wearable devices with only milliwatt level power consumption.

·Architecture innovation: In order to solve the bottleneck of the "storage wall" between chips and memory, the industry is vigorously developing 3D stacking (WoW) and in memory computing technology, greatly improving data transmission efficiency.