Comments by "MilesBellas" (@MilesBellas) on "Nvidia CUDA in 100 Seconds" video.

  1. "DeepSeek does not use CUDA because its engineers opted for PTX (Parallel Thread Execution), a low-level instruction set for Nvidia GPUs, to achieve fine-grained optimizations that CUDA cannot provide. This decision was driven by the need to overcome bandwidth limitations of the H800 GPUs used in training their models. By working at this lower level, DeepSeek reallocated GPU resources and implemented advanced optimizations, such as thread/warp-level adjustments, enabling superior performance and efficiency"
    5