And you know Nvidia can't be constent with one format for FLOPs within a single graph, 1,000,000x faster but comparing FP32 to FP8 or NVFP4 and acting like it's the same.