对话季宇：大模型非必须在GPU跑，CPU内存带宽已足够

Core Viewpoint - The conversation highlights the innovative approach of the company in utilizing CPU memory bandwidth for large model deployment, challenging the traditional reliance on GPUs for such tasks [4][8][12]. Group 1: Company Overview - The company, founded by a former Huawei expert, focuses on developing self-researched GPUs and integrated computing devices known as DeepSeek [1][4]. - The DeepSeek integrated machine, referred to as "褐蚁" (Brown Ant), is designed to be a cost-effective solution for deploying large models, with a target price significantly lower than traditional setups [5][18]. Group 2: Technology Insights - The company argues that modern server-grade CPUs can achieve memory bandwidth exceeding that of high-end GPUs, making them suitable for running large models [10][18]. - The proposed architecture aims to leverage CPU memory capabilities, which are cheaper and more efficient than traditional GPU setups, potentially reducing costs from millions to hundreds of thousands [6][18]. Group 3: Market Positioning - The company seeks to democratize access to high-performance computing by lowering the cost barrier, allowing smaller teams to engage in AI development [23]. - The strategy involves creating a product that can compete with supercomputers at a consumer electronics price point, thus fostering broader industry adoption [17][23]. Group 4: Competitive Landscape - The founder emphasizes that simply replicating NVIDIA's approach will not lead to success; instead, innovation in design and application is crucial [15][21]. - The company aims to differentiate itself by focusing on optimizing software to fully utilize CPU memory bandwidth, challenging the industry's conventional wisdom [19][22].