north-numerical-computing / tensor-cores-numerical-behavior Goto Github PK

Dear author,

I was reviewing the numerical computations in your scripts tc_test_numerics-A100-bf16.cu and tc_test_numerics-T4-A100-binary16.cu. I noticed that you've set the initial values of the first row in matrix A as 0.5 and the first column of matrix B as $2^{-24}$.

This setup results in the calculation $2^{-1}*2^{-24} + 2^{-1}*2^{-24} + 2^{-1}*2^{-24} +1 + 2^{-24}$, which indeed aligns with the conclusion in your paper.

However, when I modify the initial values of the first row in A to 1, as suggested in the original paper, the output yields 3f800002. It seems to suggest that the accurate result is obtained and that the summation doesn't necessarily start with the largest number and not involve 4 rounding errors.

Could you please confirm if my understanding is correct? Any insights you can share would be greatly appreciated.

This looks like a scoping error