<div dir="auto"><div><div class="gmail_quote"><div dir="ltr" class="gmail_attr">On Mon, Jul 15, 2019, 5:15 PM Hal Murray via devel <<a href="mailto:devel@ntpsec.org">devel@ntpsec.org</a>> wrote:<br></div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><br>
<a href="mailto:tenterline@gmail.com" target="_blank" rel="noreferrer">tenterline@gmail.com</a> said:<br>
> I come from a scientific background, where we compare results somewhat as<br>
> analog values. If the test result is off the expected by 1000%, that's bad.<br>
> If it's off 1%, better. If the error is .00001%, probably within achievable<br>
> accuracy. <br>
<br>
There is a difference between running the same experiment again to get new <br>
data and running new software on old data.<br>
<br>
Are the specs and implementation for IEEE floating point tight enough so that <br>
I should get the exact same result if I run a test on a different CPU chip? <br>
Or is there room for things like holding extra bits in temporary results so <br>
the bottom bits might be different due to round off or such?<br></blockquote></div></div><div dir="auto"><br></div><div dir="auto">Possibly. But I would say have a tight tolerance and a not so tight one. If the absolute difference is less than both paint it green, if only one then yellow, and for neither red. I'm probably wrong though.</div><div dir="auto"><div class="gmail_quote"><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
</blockquote></div></div></div>