Prime 95 torture test, different results for each core

[gamax]

Hi. I did the Prime95 2.56 torture test (the max heat and power and memory test) on my cpu (opteron 170 dual core). Basically the first test I did, the Core 0 appeared to be faster, finished the tests faster and did more tests than the Core 1. I stoped it and ran the test again, same thing, but this time the two cores worked together. Here's an SS.
Why would this happen?

http://img258.imageshack.us/img258/1112/weirdis5.jpg

PS. It appears now the Core 1 is working a bit faster (hence a bit hotter), ahead of Core 0 by 2-3 minutes or so. Is it normal that one core is faster than the other one?

 

[Andrius]

You are working in Windows.
At any given time there are somewhere between 25 and 30 processes that the OS needs to function (each has a few threads). All these threads have to execute on the two CPU cores in you system. The processing time they require varies for each core. Hence one core completes more work than the other in a given time period. This is normal. The temperature variances are also normal between cores as some tasks put less strain on the CPU. Both cores are not completly identical (minor variances in heat output).

 

[sailer]

Hi. I did the Prime95 2.56 torture test. Is it normal that one core is faster than the other one?

Yes, and if you think your results look a bit strange, you should see it on a quad core CPU. During my first run, three cores worked just fine, but one quit after three hours. Raising the CPU voltage a bit cured the one that quit, but now left the other three slightly overvolted. Nothing to worry about particularly, but something that points out to the potential problems of multiple cores and the future probable problems when Nehalem arrives and we start seeing up to 8 cores in a single CPU.

 

[Andrius]

Nothing to worry about particularly, but something that points out to the potential problems of multiple cores and the future probable problems when Nehalem arrives and we start seeing up to 8 cores in a single CPU.

I think with a quad core it already comes to the saturation point of the FSB memory access. 4 cores trying to access memory at a rate only about 1.5x faster than a single core could 3 years ago. Every core has about 40% available bandwidth of a P4 (DDR2 533MHz). Now imagine the massive overhead (and resulting performance penalties) it takes to sync 8 cores communicating through a channel that was really only meant for 1 core.
That is (amongst other good reasons) why FSB must die.

 

[kkm557]

Lol, my quad core + 8GB of RAM won't get stable at 400FSB with a P35. When I only have 2GB, I can get stable, but not with 8GB. Really annoying since I can run 400FSB on stock volts. Anyways, quad core also have some crazy temp differences. I usually have atleast a 5 degrees variance under load sometimes even as close to 10 degrees diff. As for calculation speed diff, it's what Andrius says. Maybe windows tends to favor one core over the other, but that's where affinity comes in (helpful if you have more cores than the program you are running can access).

 

[Andrius]

Lol, my quad core + 8GB of RAM won't get stable at 400FSB with a P35. When I only have 2GB, I can get stable, but not with 8GB. Really annoying since I can run 400FSB on stock volts.

That is likely a timing issue. Try at 5-5-5-18 and if you have a Gigabyte board set the "RAM Performance Enhance" to "Standard".

 

[montanabay]

How are you setting the affinity for each instance Prime95? When I set it in the Prime95 GUI each will give me uneven speed and temps, but when set them in task manager it's a dead heat.

Cheers,

Josh

 

[Andrius]

In Prime95 version 25.6 there's no need to set thread affinity.
It spawns a thread for each core by default.

Even processor time and temperature across all cores is impossible.

There are 21 system processes in a fresh XP install. Running taskmanager and Prime95 over a period of 12 hours my cores are spread over 3-4 tests (each test runs about 2 minutes).

Temperatures can be optimized by lapping the IHS and the base of the heatsink. With a quadcore a 10°C spread is quite normal.