Seems like it is more oriented for LLM inference where it has ~80% higher memory bandwidth and ~30% higher memory and tensor core performance, and 100% increase in PCIe bandwidth.