> More than twice as fast as Nvidia 4090 for AI.

Not in memory bandwidth which is all that matter for LLM inference.