This blog post describes how to build a dense B200 attention kernel from scratch in CUDA and PTX, achieving 94.4% performance on 4K, 8K, and 16K shapes used in the FA4 paper. The main contribution is a visual guide with diagrams and code that explains how to optimize the kernel's softmax operation using hardware instructions like ex2.approx.ftz.f32, resulting in improved performance about 10%.