accuracy 32 bit/64 bit atk

[Zexter]

I want to know whether running atk linux 32 bit software on a 32 bit OS would produce a different result(the places after the decimal point)when compared to running the same input in atk linux 64 bit on a 64 bit OS ......If so how to get same accuracy in 32 bit os 32 bit atk.

[Anders Blom]

The primary accuracy concern is not really between 32 or 64 bit, but rather between single and double precision. ATK uses only double precision, so actually the bit representation of all numbers are identical on 32/64. Differences in the results can however occur because the self-consistent loop is highly non-linear, and small differences in starting conditions can propagate to a larger difference in the end result. The magnitude of these differences vary strongly, from 1e-4 to 1e-10. Still, through the consistent use of double floating representation, these differences are on the level where they have no real influence on the physical relevance of the results; after all, an atom position specified to the accuracy level even of 0.001 Angstrom makes no sense, and if a transmission coefficient is 0.12345 or 0.12344 also cannot be said to be significant.

[Nordland]

In addition I might add, that if you set the tolerance to 1e-5 the difference is in the order of 1e-5. If you set the tolerance to 1e-8, the difference is order of 1e-8.