[macstl-dev] Opteron benchmarks

[Glen Low]

Hi All

While waiting for macstl 0.3.1, you might be interested in the latest  
macstl benchmarks. A quad 2.6GHz Opteron (AMD64) landed in my day job  
the other day and I had to compile macstl and run it on the beast.

First, using the old Visual Studio C++ 2003 which produces 32-bit  
code. For the key to these values, look at our last benchmarks:  
http://www.pixelglow.com/stories/macintel-faster-than-altivec/

MA    1,282    1.6x
IP    3,226    7.5x
Poly    709    1.2x
Hypot    279    3.1x
Trig    24.7    3.2x

Next I used Visual Studio C++ 2005 Beta 2 which still produces 32-bit  
code. There is a restriction on benchmarking .NET Framework code, but  
since no MSIL was harmed in this procedure, I should be in the clear.

MA    1,075    5.7x
IP    1,282    8.0x
Poly    917    4.3x
Hypot    279    4.0x
Trig    31.4    8.9x

Now for the piece de resistance. I flicked on 64-bit code aka x64 aka  
x86-64 aka EM64T. The main benefit for macstl will be the additional  
registers over the register-starved 32-bit ISA.

MA    2,128    3.3x
IP    -- crash --
Poly    1,282    2.6x
Hypot    291    5.1x
Trig    42.7    9.8x

The multiply-add was 98% faster and the trig test was 36% faster than  
over 32-bit code.

Versus the 3.4GHz Xeon, the Opteron draw even on the trig test while  
it was 3x faster on the multiply-add.

Versus the 2.0GHz G5, the Operton was about 60% of the trig test  
speed while it almost drew even on the multiply-add.

Cheers, Glen Low


---
pixelglow software | simply brilliant stuff
www.pixelglow.com
aim: pixglen

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://www.pixelglow.com/lists/archive/macstl-dev/attachments/20050710/76781cc1/attachment.html