|
|
|
|
|
|
by Swapnopam
9 days ago
|
|
|
Agreed with you on that. One of the reasons for my post was, I was told by some engineers over at aeronautical companies that such float drift is trivial — anything beyond 10^-5 or 10^-8 becomes meaningless. Nearly the same was said by some engineers at HFT and quant firms. Which led me to wonder if cross-hardware drift is really a big deal. From what I've tested, even differences between an Intel i5 and Intel i7 pile up, let alone between Intel and AMD. Curious if the "it's within tolerance" answer holds in your RF work or if you've hit cases where the hardware boundary actually mattered. |
|
|
i5 and i7 are terms which lack precision. Are you seeing different results for the same operations on the same arguments in the same order across an i5 and an i7 with the same core architecture?
Regardless, it would be great if you have examples of operations that have different results on different processors; my expectation was that I could expect same width, same operation, same arguments to provide the same result across most x86 processors (early pentiums not dividing properly and maybe more variance in the socket 7 days where there were so many vendors). Most of the issues I've seen documented were not running the same sequence of operations especially around 80-bit floats on x87 if something spills to a 64-bit double in memory. All that said, I prefer to deal in integers whenever possible, I know the edges of floating point are sharp, and I'd rather not be cut.