Core Viewpoint - Domestic AI chip companies like Sunrise are focusing on the inference chip market, differentiating themselves from competitors like Nvidia by targeting specific market segments rather than attempting to cover both training and inference simultaneously [2][4]. Company Overview - Sunrise, spun off from SenseTime's chip division, aims to establish itself in the inference chip market, having completed its first round of external financing by the end of 2024 and raised nearly 1 billion yuan in July 2023 [2][3]. - The company is led by Xu Bing, co-founder of SenseTime, and has a management team with backgrounds from Baidu [2]. Product Development - Sunrise has launched three generations of inference chips: - The first-generation S1 chip, launched in 2020, focuses on visual inference and has sold over 20,000 units [3]. - The second-generation S2 chip, set to begin production in September 2024, claims to achieve performance close to 80% of Nvidia's A100 [3]. - The third-generation S3 chip is expected to be officially launched in May 2025, optimized for large model inference and supporting low-precision data formats [3]. Market Trends - The demand for inference computing power is rising due to the accelerated adoption of AI applications, prompting Sunrise to focus on this segment [4]. - The industry is witnessing a shift towards high-performance inference chips, as the market for high-performance training chips is perceived to be limited [4]. Strategic Partnerships - To reduce customer migration costs, Sunrise has chosen to be compatible with Nvidia's CUDA parallel computing framework, facilitating easier adoption for developers [5]. - The company has established partnerships with various industry players, including SANY Group, Fourth Paradigm, Midea Group, and others, ensuring customer engagement from the design phase [5]. Design Considerations - Achieving a balance between computing power and memory bandwidth is crucial for optimizing the cost-performance ratio of inference chips [5]. - Sunrise emphasizes the importance of aligning chip design with target computing tasks to avoid inefficiencies that could lower the chip's value proposition [5].
商汤分拆的AI芯片公司，为何全盘押注模型推理市场？