“Groq recommends running the prefill stage of LLM inference exclusively on GPUs because they are highly efficient at that task.”