TF32 is 19 bits, not 18 bits. There's an additional bit for sign. https://blogs.nvidia.com/blog/2020/05/14/tensorfloat-32-precision-format/ https://blogs.nvidi

TF32 is 19 bits, not 18 bits. There's an additional bit for sign. https://blogs.nvidia.com/blog/2020/05/14/tensorfloat-32-precision-format/ https://blogs.nvidia.com/blog/2020/05/14/tensorfloat-32-prec...