电子行业专题报告：TPU：为更专用的AI计算而生

Investment Rating - The report maintains a rating of "Outperform" for the electronic industry, specifically highlighting the TPU sector as a key area of interest [3]. Core Insights - The TPU (Tensor Processing Unit) is positioned as a leading AI chip, designed specifically for AI computations, showcasing significant advantages over traditional GPGPU architectures in terms of performance and power efficiency [1][7]. - Google has established a strong foothold in the TPU market, with an estimated shipment of over 2 million TPU units by 2023, primarily utilized for internal operations and a portion available for external leasing [34][35]. - Domestic TPU manufacturers, particularly Zhonghao Xinying, are emerging as competitive players, achieving profitability and securing significant orders, indicating a growing market for domestic AI chips [2][41]. Summary by Sections 1. TPU Development and Advantages - TPU was developed to address the inefficiencies of traditional CPUs and GPUs in handling deep learning workloads, leading to the first TPU release in 2015 [7]. - The TPU architecture is optimized for matrix multiplication, making it highly efficient for AI algorithms, with a notable performance-to-power ratio compared to GPGPU [1][19]. - The TPU's unique design allows for high throughput in matrix operations, leveraging a pulsed array architecture that enhances parallel computation capabilities [20][23]. 2. Google’s TPU Ecosystem - Google’s TPU ecosystem has grown significantly, with a focus on internal applications and a leasing model for external clients, including major AI startups [34][36]. - The TPU's integration with TensorFlow enhances its performance, allowing for efficient processing of complex machine learning models [33]. 3. Domestic TPU Manufacturers - Zhonghao Xinying has successfully launched its first TPU chip, achieving significant performance metrics that surpass NVIDIA's A100, while also maintaining a lower cost per unit of computing power [41][42]. - The company reported revenues of 485 million yuan and a net profit of 80 million yuan in 2023, indicating strong self-sustainability and growth potential [43].