B200 Attention Kernel from Scratch to Near-SOTA in 60 Diagrams

A developer built a B200 attention kernel from scratch in CUDA, achieving 94.4% of FlashAttention-4 performance through a visual learning progression of 60 diagrams that incrementally add optimizations. The blog serves as a beginner-friendly foundation for GPU kernel research on Blackwell hardware, with all code available and a capstone project demonstrating the kernel's application in video generation.