Comment by throwawaymaths

nah, it was designed for hpc and raw flops. llm inference really requires memory bandwidth.