Tim,
On 2005 Aug 24 , at 20.04, Tim Jenness wrote:
> My timings are
>
> Apele (2.7 GHz PPC 8GB):
> no opt 36 s
> -O2 30
> -O3 30
> -O3 -faltivec 30
> -O3 -faltivec -mcpu=powerpc64 24
> -O3 -faltivec -mcpu=powerpc64 -m64 12 *
The timings on my laptop are:
-O0
real 4m21.834s
user 1m27.934s
sys 0m14.639s
-O2
real 1m38.204s
user 0m58.125s
sys 0m6.467s
-O3
real 1m5.366s
user 0m56.284s
sys 0m2.975s
-O3 -faltivec
real 1m3.961s
user 0m56.280s
sys 0m2.927s
-O3 -maltivec -mcpu=powerpc
real 2m21.231s
user 0m58.241s
sys 0m6.582s
This is a 1.5GHz G4 with 1GB of memory, and appears to be roughly
half the speed of Tim's 2.7GHz G5 (which is odd). Since it's a G4,
the powerpc64 options don't work; it's nonetheless bizarre that
adding -mcpu=powerpc actually slows things down, and matches just the
-O2 one (am I misunderstanding what G4s are???). This is gcc 4.0.0.
Is this actually telling us that the PowerPC memory bandwidth is
poor? Or that gcc 3.x is ill-optimised for it. I'm hopeless at
interpreting benchmarks.
These are with the system unloaded. Apart from the -O0 one they're
faster than or comparable with Steve's G5 -- since this is testing
memory bandwidth, it doesn't make much sense to do it with other
things swapping in and out, as Steve noted.
> Looking on this page, the FPU of the G5 appears to be good, but the
> other aspects are poor:
>
> http://www.tux.org/~mayer/linux/results2.html
That might be overstating it.... That table appears to be ranked by
INT performance (I thought PowerPCs were supposed to be wonderful for
integer stuff).
See you,
Norman
--
----------------------------------------------------------------------
Norman Gray / http://www.astro.gla.ac.uk/users/norman/
Physics & Astronomy, Glasgow University, UK
|