Result Accuracy

[Little-six-Wang]

Dear, author.
I would like to ask what range of values for $\frac{||Ax-b||}{||b||}$ can be considered acceptable for verifying the correctness of the program?
During my tests, I found that for different input matrices, the error can sometimes reach the magnitude of10 . Additionally, different kernels used in the numeric factorisation phase also produce varying errors—some are very small, close to 0, while others result in errors on the order of 10.
Looking forward to your reply.

[edwardnjust]

I met the same problem. When I run the program on RTX 5880, the error is $10^{16}$, but I run it on RTX 4090, the error is $10^{-16}$. Maybe it related to the machine architecture? I guess. -gencode=arch=compute_80,code=sm_80 in my make.inc