Nvidia says it can shrink LLM memory 20x without changing model weights - 信息索引