An Investigation of FP8 Across Accelerators for LLM Inference