What happens when you run a CUDA kernel
Summary
This article provides an in-depth tour of CUDA kernel execution, tracing the path from host-side nvcc-compiled code to device-side SASS, including PTX translation, QMD setup, pushbuffers, doorbells, and the GPU's warp scheduling and memory-path behavior. It uses the example kernel vadd and ties together concepts like memory coalescing, L1/L2 caches, and DRAM bandwidth with practical launch details.