Re: confused about performance
On Thu, Jun 14, 2007 at 05:00:00PM +0200, Leopold Palomo-Avellaneda wrote:
> the question is how many apps need this quantity of ram? Also, the mobo have
> to support that. My old mobo I think that it's limited to 4Gb.
It doesn't have to be many. All it takes is one for it to be worth
using 64bit for someone.
> with this are you saying that all the apps that do some kind of operation with
> doubles and floats will be better?
Yes. Of course it is possible to tell gcc to use sse for floating point
even on 32bit, but it isn't the default there so libraries would
normally not be compiled that way. x87 was a crappy design that only
made any sense at all back when the fpu was a seperate chip entirely.
Note however that you loose 80bit floating point with sse (it only does
64bit like just about every other system ever made. 80bit was a weird
intel x86 thing it seems).
> sure, but I don't think that you have a loss, that I think that it's difficult
> to me to undertand.
Based on what I have seen, the extra registers 64bit gets to use more
than makes up for the extra cache/ram space needed by larger points, so
for most things you get a small performance gain, some things gain more
(gzip's 20% improvement is pretty nice, and could be entirely due to the
extra registers). I think the cases where 32bit beats 64bit are very
very rare, and probably not worth worrying about. Well if you are
running on a Pentium 4, 32bit tends to beat 64bit most of the time, but
it had a design where 64bit seemed like a tacked on hack. The Athlon 64
and Core 2 on the other hand seem to really like 64bit code.
> it doesn't worsewhile but thanks for the example, is very appropriate.
gzip running 20% faster seems like a reason for 64bit to me. :) Simpler
memory map seems worth it. 32bit is rather mature, 64bit is still new.
That means 64bit is most likely to improve over time, while 32bit may
not get much better than it already is.