The LLM model performance is a sideshow. They are demoing linear scaling on their "cluster on a wafer" tech (a vertically scaled single chip "the largest comput

The LLM model performance is a sideshow. They are demoing linear scaling on their "cluster on a wafer" tech (a vertically scaled single chip "the largest computer chip ever built and the fastest AI processor on Earth") vs distributed training. They have a cloud service offering for example fine tuning and pretraining well known LLM models on their chips. (est. $40k to train GPT3 6.7B in ~11 days)