“In a hybrid GPU-LPU system for LLM inference, Groq's recommended architecture runs the projection layers on its LPUs and the attention mechanism on the GPUs.”