智通财经网·2025-08-24 07:48Core Viewpoint - DeepSeek's announcement of its new model DeepSeek-V3.1 utilizing UE8M0 FP8 Scale parameter precision has sparked significant interest in the capital market, leading to a surge in stock prices of chip companies like Cambrian. However, industry insiders express a more cautious outlook regarding the practical value and challenges of FP8 in model training and inference [1][4]. Group 1: DeepSeek's Impact on Capital Market - The launch of DeepSeek-V3.1 has led to a strong reaction in the capital market, with stock prices of chip companies rising sharply [1]. - The industry response at the 2025 Computing Power Conference was more subdued, focusing on the actual value and challenges of FP8 rather than the excitement seen in the capital market [1]. Group 2: Understanding FP8 - FP8 is a lower precision format that reduces data width to 8 bits, enhancing computational efficiency compared to previous formats like FP32 and FP16 [2]. - The direct advantages of FP8 include doubling computational efficiency and reducing network bandwidth requirements during training and inference, allowing for larger models to be trained or shorter training times under the same power consumption [2]. Group 3: Limitations of FP8 - While FP8 offers speed advantages, it can lead to calculation errors due to its limited numerical range, necessitating a mixed precision training approach to balance efficiency and accuracy [3]. - Different calculations have varying precision requirements, with some operations being more tolerant of lower precision [3]. Group 4: Future of DeepSeek and FP8 Standards - DeepSeek's use of FP8 is seen as a signal that domestic AI chips are entering a new phase, providing opportunities for local computing power manufacturers [4]. - The industry acknowledges that while FP8 represents a step towards computational optimization, it is not a panacea, and the actual implementation results are crucial [4]. - The transition to FP8 may require an upgrade across the entire domestic computing ecosystem, including chips, frameworks, and applications [4]. Group 5: Challenges in Large Model Training - The core bottlenecks in large model training and inference include not only computational scale but also energy consumption, stability, and cluster utilization [5]. - There is a need for advancements from simple hardware stacking to more efficient single-card performance and optimized cluster scheduling to meet growing demands [5].