Linus Torvalds (torvalds@linux-foundation.org) on 1/3/11 wrote:
---------------------------
>Carlie Coats (coats@baronams.com) on 1/3/11 wrote:
>>
>>And all this apparently with "least-common-denominator" executables
>>that are not compiled for Sandy Bridge, and presumably cannot take
>>advantage of its floating point improvements.
>
>Well, the biggest improvement is likely just the memory
>pipeline, so even without actually using AVX, I do think
>those tests are really taking advantage of sandybridge.
>
>Even with old binaries, you'll be doing more loads per
>cycle etc.
>
>Sure, there's probably more to be had, but I think it's
>very healthy to primarily care about older binaries. If you
>need to recompile to get the advantage of a new uarch,
>there's something seriously wrong with the hardware.

I should have said that my point of view is that of
an environmental modeler/forecaster -- I *always* have
to compile the models from scratch, and every extra
FLOP I can squeeze out shows up as better (either more
timely or more detailed) forecast. It's a rather narrow
POV, I'll admit...

[You don't want to get me started on how much trouble
the shared-library ideology causes someone like me,
with a couple of hundred clients -- who are running
half-a-hundred different combinations of Linux
distributions! (There's _no_way_ I can distribute
anything *but* source.)]