Is it feasible to run LLM inference comparably without CUDA or Rocm? How much of the cost performance goes away?