If there are in fact stability issues, I wonder which is cheaper: using this fancy algorithm or changing bfloat16 to something like bfloat14 and using a more st

If there are in fact stability issues, I wonder which is cheaper: using this fancy algorithm or changing bfloat16 to something like bfloat14 and using a more stable matmul.