“In processing an LLM token, GPUs are better suited for compute-constrained matrix multiplications, while LPUs are better for memory throughput-constrained ones.”