4 ms·

I wish they had mandated the algorithms used in the distributions. Even with the same generator initialized with the same seed, the normal distribution, for exa

I wish they had mandated the algorithms used in the distributions. Even with the same generator initialized with the same seed, the normal distribution, for example, will give different results across platforms. At least the results of the generators are consistent across platforms.

Someone 11y ago
Is that feasible for distributions over the reals across CPU architectures or maybe even revisions? FPUs typically do not compute the best possible result for floating point operations. Also, instructions like 'estimate one over square root' have multiple 'correct' answers. Edit: https://www-01.ibm.com/support/knowledgecenter/#!/ssw_aix_71/com.ibm.aix.alangref/idalangref_frsqrte_instrs.htm https://www-01.ibm.com/support/knowledgecenter/#!/ssw_aix_71... shows things are worse: "The estimate placed into register FRT is correct to a precision of one part in 32 of the reciprocal of the square root of FRB. The value placed in FRT may vary between implementations and between different executions on the same implementation."
- api_or_ipa 11y ago
  Is this an indication of a randomly seeded Newtown's method?
  - Someone 11y ago
    But why would they? My guess would be that they don't bother to clear a carry flag or are computing the estimate using a few extra significant bits, and do not bother to initialize those at start, but that's 100% guess. I know very little about designing CPUs.