Nvidia’s new Blackwell chips trained Meta’s large Llama 3.1 model in just 27 minutes. They used fewer GPUs and were over ...