ParentFull threadmy123·The 30x number is for a really narrow scenario tbh. Running a GPT 1.8T parameters (w/ MOE) on one GB200View on HN