DeepSeek AI has introduced several advancements in matrix multiplication and training algorithms aimed at enhancing efficiency in artificial intelligence applications. The newly released DeepGEMM is a lightweight CUDA library that accelerates FP8 General Matrix Multiplications (GEMMs) on NVIDIA Hopper GPUs, achieving speeds up to 2.7 times faster. Additionally, DeepSeek has launched DualPipe, a bidirectional pipeline parallelism algorithm designed for computation-communication overlap in V3/R1 training. This algorithm aims to improve the efficiency of AI training processes. Another notable release is SkipPipe, a communication-efficient, pipeline-parallel training method that can reduce distributed training time by up to 55% and is scalable to theoretically infinite model sizes. These innovations reflect DeepSeek's commitment to pushing the boundaries of parallelism and optimization in AI training.