has anyone numbers on how many queries/second in inference time chatGPT can serve?