Hand-written CUDA Tensor Core GEMM kernels on Blackwell (sm_120) and Hopper (sm_90) — raw mma.sync reaching 106% of the cuBLAS-TC kernel on sm_120, CUTLASS 3.x wgmma at 85.5% of nvjet on H100, and an FP16→FP8→MXFP4 precision ladder. Every number reproducible from committed bench data.