Yes it is slow for training/fine-tuning decently sized LLMs and unusable for diffusion models. What are you training were you get usable speeds?