One multiply, one RTX 5090, eight kernels
One small-batch matrix multiply on a current NVIDIA GPU, followed from 748 microseconds to 29 across eight measured CUDA kernels, with the code, the machine instructions and the profiler counters that changed at each step.
