Comment by wolfgangK

For LLM inference, I don't think the PCIe bandwidth matters much and a GPU could improve greatly the prompt processing speed.