Workflow
国泰海通:ROCm生态或加速完善 海光中标数十亿订单
智通财经网·2025-10-10 06:01

Core Insights - AMD has signed a 6GW computing power chip cooperation agreement with OpenAI, which is expected to accelerate the ROCm ecosystem development [1] - The Haiguang DTK software stack is fully compatible with the ROCm ecosystem, potentially reducing the adaptation costs for large models in China [2] - Industrial and commercial banks, along with China Unicom, are procuring Haiguang chips worth several billion yuan, indicating a rapid penetration of Haiguang chips in the domestic market [3] Group 1 - AMD and OpenAI's agreement includes the deployment of a total of 6GW of AMD GPU computing power, starting with 1GW based on the MI450 series GPU in the second half of 2026 [1] - The collaboration is seen as a response to the significant gap in AI computing power supply and aims to enhance the synergy between AMD and OpenAI in both hardware and software [1] - ROCm 7 is expected to be available by Q3 2025, offering over 3.5 times the inference capability and three times the training capability compared to ROCm 6 [1] Group 2 - The Haiguang DCU features a complete software stack, including DTK, development toolchain, and model repository, which supports mainstream deep learning frameworks like TensorFlow and Pytorch [2] - This compatibility is anticipated to effectively lower the costs associated with adapting large models domestically [2] Group 3 - Industrial and Commercial Bank of China has announced a procurement project for Haiguang chip servers, with a potential total procurement amount reaching 3 billion yuan [3] - China Unicom's procurement project for Haiguang servers has a budget of 2.782 billion yuan, indicating strong demand for domestic chip solutions [3] - The financial and telecommunications sectors are expected to lead the way in achieving comprehensive domestic production, thereby accelerating the penetration of Haiguang chips [3]