Xuan Gu Bao·2025-12-16 15:17Core Insights - The article highlights a significant breakthrough in cross-architecture large model inference technology verification, achieved by the China Telecom Research Institute in collaboration with various industry partners [1] - This technology enables seamless operation of the same operator source code across three types of chips: Nvidia, Ascend, and Muxi, addressing the challenge of multi-architecture adaptation [1] - The innovation aims to enhance the usability of domestic computing power, facilitating the transition from "usable" to "user-friendly" and "easy to use," thereby supporting the autonomous and diversified development of computing infrastructure in China [1] Industry Summary - The successful implementation of a heterogeneous large model inference framework based on Triton has compressed the operator adaptation cycle from "weekly" to "daily," achieving performance levels of 90% compared to native operator libraries [1] - This advancement is expected to provide business entities with intuitive and precise chip selection decision support, which is crucial for the commercialization of large models [1] - The article mentions related A-share concept stocks such as Tuosida and Mengwang Technology, indicating potential investment opportunities in this sector [1]