PT9 Floating Point Math score difference between 1005 and 1014

[TMoney128]

I recently noticed the CPU test for Floating Point Math is about 1000 points lower if I compare build 1005 and 1014. My Xeon X5650 scores about 9500 in b1005 and only 8500 in b1014. I ask because I don't see anything in the change logs between the two builds that could've effected the scores. I also confirmed there is no other influences on the scores by running both builds simultaneously. Thanks.

[David (PassMark)]

Nothing has changed in the tests.

There was this long thread about really really poor performance in the floating point test. But your score doesn't seem nearly bad enough to qualify as the same issue.

Can you run each build a few times and take the max value. Random background tasks on windows can give the odd low result, so doing multiple runs and taking the max can help avoid this.

[TMoney128]

OK I ran each version 5 times on default Medium length.

B1005 - 9443
B1014 - 8471

I also uninstalled, re-installed, installed on a different drive, found no change on 1014.

I changed the number of processes for the CPU tests in both builds on my 6 core/12 thread Xeon X5650:

B1014 / B1005
2 - 3297 / 3298
4 - 6588 / 6596
6 - 9815 / 9826
8 - 9326 / 9703
10 - 8757 / 9544
12 - 8465 / 9449

They both look about the same until you get into the hyperthreaded cores. Both versions taper off after the 6 physical cores but 1014 loses almost 1400 points vs. only 400 points or so on 1005. Is there any way I can download the other builds between 1005 and 1014? That way I can tell you exactly what version the change happened since maybe it was something unintentional. Thanks for the help.

[TMoney128]

I would also like to add I tested the FP Math on the 32 bit version of build 1014 vs. the 64 bit version:

32bit - 9143
64bit - 8486

Oddly enough the 32 bit version scores quite a bit higher. Almost there!

[David (PassMark)]

I think the fact that the scores are pretty much the same between 2 and 6 cores shows there was no change to the test itself.

Also slightly surprising that hyperthreading gives a negative performance increase. But this is not unheard of.

Maybe there is some change in the compiled executable, causing it to use a different part of the memory address space, which in turn causes caching issues on the CPU. This is just speculation, we don't really know.

You can't really compare the 32bit and 64bit releases. There should be a difference between these.

Here are a couple of temporary links to intermediate builds
www.passmark.com/temp/petst9.0.1011.exe
www.passmark.com/temp/petst9.0.1009.exe
www.passmark.com/temp/petst9.0.1006.exe

[TMoney128]

Thanks for the links. I retested with 5 iterations on each keeping the best result:

B1006 - 9437
B1009 - 8475
B1011 - 8484

So it looks like the change happened somewhere between 1006 and 1009. If you have 1007 and 1008 I can test them too. Thanks for taking the time looking into this.