科大讯飞：讯飞星火对标A100的训练效率优化后达到85%-95%以上

Core Viewpoint - The company has made significant investments in optimizing the training and inference cost efficiency of its large models under limited computing resources, achieving substantial improvements in performance metrics compared to industry standards [1] Group 1: Technological Advancements - Since May 2023, the company has collaborated with Huawei to overcome various technical challenges, including high-speed interconnection, hidden computation communication, and optimization of training and inference efficiency [1] - The training efficiency of general large models and deep inference models has improved from an initial 30%-50% to over 85%-95% when benchmarked against NVIDIA's A100 [1] Group 2: Breakthroughs in Domestic Computing Power - In 2025, the company achieved significant breakthroughs in two areas: enhancing the training efficiency of long-thought chain reinforcement learning from 30% to over 84% against the A800 benchmark, and improving the full-link training efficiency of MoE models from 30% to 93% [1] - These advancements represent a major leap from 0 to 1 in the domestic computing power sector, indicating a strong potential for further cost reductions in training [1]