计算机行业定期报告：Deepseek发布全新注意力机制NSA

Investment Rating - The industry rating is "Outperform the Market" [6][35]. Core Insights - The NSA (Sparse Attention Mechanism) proposed by Deepseek, Peking University, and the University of Washington aims to address performance bottlenecks in traditional attention mechanisms for long contexts and multi-turn dialogues. It features a three-parallel branch architecture (Token Compression, Token Selection, Sliding Window) combined with a learnable gating mechanism to dynamically balance global and local attention, significantly improving inference speed while maintaining accuracy [3][12][13]. - The hardware optimization is based on the Triton framework, enhancing memory access efficiency through shared KV data, high-bandwidth HBM, and on-chip SRAM collaboration, making it suitable for large language model acceleration and long document understanding [3][14][19]. AI Data Update - In the overseas market, from February 14 to February 20, 2025, the download volume of ChatGPT has gradually rebounded, while Gemini, Perplexity, and Claude have remained stable. ChatGPT and Gemini are the top two AI applications in terms of downloads since the beginning of 2024 [20][21]. - In the domestic market, during the same period, Deepseek's download volume has slightly decreased, while Kimi, Tongyi, Xinghuo, and Wenxin Yiyan have remained stable. Notably, Tencent Yuanbao's integration with Deepseek has led to a significant increase in downloads, now exceeding 300,000 times per day [21].