大语言模型算力度量模型  

Computational Measurement Model for Large Language Models

作　　者：刘永生[1] 张岩[1] 周广[1] 曹畅[1] Liu Yongsheng;Zhang Yan;Zhou Guang;Cao Chang(China Unicom Research Institute,Beijing 100048,China)

机构地区：[1]中国联通研究院,北京100048

出　　处：《邮电设计技术》2024年第9期20-23,共4页Designing Techniques of Posts and Telecommunications

摘　　要：面对大语言模型对算力需求的快速增长,传统的摩尔定律已经难以满足需求,而大语言模型的扩展法则表明更多参数、更多数据和更多算力能够得到更好的模型智能。针对大语言模型的算力度量问题开展研究,旨在评估大语言模型的算力需求。提出大语言模型训练的算力度量模型和大语言模型推理的算力度量模型,并通过理论分析提出了相应的计算方法。In the face of the rapidly increasing demand for computing power in large language models,traditional Moore's Law is no longer sufficient to meet the demand,while the expansion rules of large language models indicate that more parameters,more data,and more computing power can lead to better model intelligence.Research is conducted on the measurement of computing power for large language models in order to evaluate the computing power requirements of large language models.It proposes a computational power measurement model for training large language models and a computational power measurement model for inference of large language models,and the corresponding calculation methods is put forward through theoretical analysis.

关 键 词：大语言模型 算力度量 人工智能 

分 类 号：TN915.5[电子电信—通信与信息系统]