How CUDA spreads a matrix multiply

How does CUDA spread a language model's matrix multiplication across thousands of GPU workers?