I would never have guessed that my laptop would beat them all.

Does this mean that Linux x32 is the fastest OS in the world, and that 32-bit is faster than 64-bit, and that clockspeed means everything, and dual cores are better than quad and single cores?

I want to find out what is the fastest linear and parallel processing system on the market, and I really hope M5 will be it.

System ID	M1	M2	M3	M4	M5	M6	M7	M8
System Name	IBM System p5 505	IBM System x 3650	Lenovo ThinkPad T61p	Siipi Blackhawk	IBM System p 570	IBM System x 346	Siipi Falcon	Siipi Osprey
System Specifications	POWER5 1.9GHz	2x Intel Xeon Quad Core 2.66GHz	Intel Dual Core 2.66GHz	Intel Quad Core 2.4GHz	4x POWER6 4.7GHz	2x Intel Xeon 2.4GHz	Intel Pentium 4 3.2GHz	Intel Dual Core 1.8GHz
Operating System	AIX 5L 5.3	Windows 2003 Enterprise R2 x64	Windows XP SP2 x32	Windows XP SP2 x32 / Linux Ubuntu 7.10 x32	AIX 6.1	Windows 2003 Enterprise x32	Windows 2003 Enterprise x32	Windows XP SP2 x32
Speedtest 1.0 32-bit	-	25.125s	26.640s (14.312s+12.328s)	27.781s (Windows) 48.010s (Linux, default) 21.590s (Linux, optimized 16.660s+4.930s)	-	58.406s	45.312s	38.296s
Speedtest 1.0 64-bit	199.391s (default) 51.450s (optimized)	122.937s	-	-	coming in 2008-03	-	-	-

Maybe I should explain a bit further why these systems were used in the test:
M1: the cheapest System p you can get, we wanted to have a test server for evaluating the production server (possibly JS22 or 570)
M2: our currently fastest System x server
M3: my office laptop
M4: my home gaming PC
M5: an vision how our first production System p could look
M6: our old production server
M7: a low cost server (high GHz single HyperThreading core (=semi-dual core))
M8: a low cost gaming PC (low GHz dual core)

Speedtest 1.0 source code (all-bit, all platforms) is here:

#include "stdio.h" #include "time.h" class testclass {    public:    int x;    testclass(void) {        x=1;    }    ~testclass(void) {       x=0;    } }; int main(int argc, char **argv) {    double n=0;    long i=0;    long t1=0;    long t2=0;    long t1t=0;    long t2t=0;    t1t=clock();    printf("Speedtest 1.0 (c) 2008 Siipi\n");    printf("Counting 10 billion floating points...\n");    t1=clock();    while(n<100000.0)    {       n+=0.00001;       i++;    }    t2=clock();    printf("Done. i=%ld, n=%f, time=%fs.\n",i,n,    (double)(t2-t1)/CLOCKS_PER_SEC);    printf("Creating and deleting 1 billion class objects...\n");    t1=clock();    i=0;    while(i<100000000)    {       testclass *a=new testclass();       delete(a);       i++;    }    t2=clock();    printf("Done. i=%ld, time=%fs.\n",i,    (double)(t2-t1)/CLOCKS_PER_SEC);    t2t=clock();    printf("Total time=%fs.\n",(double)(t2t-t1t)/CLOCKS_PER_SEC);    return(0); }

To compile and run Speedtest 1.0 under AIX/Linux/MacOSX/Sun/Cray/IRIX/AmigaOS, enter the following commands:
g++ st.cpp
time ./a.out

To compile and run Speedtest 1.0 under Windows, enter the following commands:
(Launch Visual Studio 2005 C++, select: File/New/Project From Existing Code/Console Application)
(Select Release instead of Debug)
(Choose Build/Build Solution)
\programs\rktools\ntimer release\st.exe (ntimer.exe comes with the Windows 2003 Resource Kit Tools)

You don't necessarily need to run Speedtest 1.0 with time/ntimer, I just used it to double-check that my time measurement was working fine.

The program output will look like this:
Speedtest 1.0 (c) 2008 Siipi
Counting 10 billion floating points...
Done. i=1410063201, n=100000.000003, time=15.078000s.
Creating and deleting 1 billion class objects...
Done. i=100000000, time=12.703000s.
Total time=27.781000s.

Unfortunately I don't have all systems which exist, but here are some precompiled binaries:

Windows 32-bit	Windows 64-bit	AIX 64-bit	Linux 32-bit
ftp://ftp.siipi.com/st.exe	ftp://ftp.siipi.com/st64.exe	ftp://ftp.siipi.com/st64	ftp://ftp.siipi.com/st

The source code can be also downloaded, to make it a bit easier on systems which have no web browser installed: ftp://ftp.siipi.com/st.cpp.

I would be glad to know if someone's server/workstation can run this under 21s.

I managed to get the POWER5 code twice faster with the following compiler options:
g++ -O3 -fomit-frame-pointer -fstrict-aliasing -mcpu=power5 st.cpp

The Linux code went also more than twice faster with these options:
g++ -O3 -mtune=pentium4 st.cpp

Furthermore, the AIX programs might also run faster when compiled with IBM VisualAge AIX 6.0 C++ (especially with the XL C/C++ addon) or Intel C++, although Linux GNU C++ beats Microsoft Visual Studio 2005 C++ in optimized machine code speed, which also surprised me, let's say vastly.