[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: Itanium2@900MHz slower than alpha@666MHz ?

>>>>> On Tue, 21 Oct 2003 01:57:47 +0200, Ionut Georgescu <george@physik.tu-cottbus.de> said:

  Ionut> Thanks, I'll forward the message to that list. Can't check for dmesg
  Ionut> right now because I started some benchmarks from www.fftw.org and I
  Ionut> don't want to 'disturb'. I mainly want to see if I can reproduce their
  Ionut> results.

Sounds like a good course of action.

  Ionut> However, this is well tested code and it's been running on
  Ionut> the alpha for months. Had there been any FP error, it would
  Ionut> have crashed on the spot.

I wasn't very clear: fpswa stands for "floating-point software assist"
which is a trap that gets invoked when encountering a corner-case that
isn't handled in hardware (such as operations on infinity, NaN, or
denormals).  The traps are handled in the kernel and do _not_ result
in an error, but such operatiosn of course to run more slowly.  Alpha
behaves very similar, so like I mentioned earlier, it's unlikely to be
an issue with fftw2.

There are a lot of numbers at fftw.org, but from a cursory look at it,
it does seem like peak performance of a 900MHz Itanium 2 is
significantly better than peak performance of an 833MHz Alpha.

Oh, one thing I noticed: the folks at fftw.org used "ecc -O3" whereas
you used "ecc -O2".  The main difference with -O3 is that it turns on
data prefetching.  This can make a huge difference when dealing with
relatively large data sets.


Reply to: