What's the problem whit the cpu? Or is it cpu error anyway?
Given the minimal information you've provided (what mother board do use, is Turbo enabled and what is the multiplier set to, what CPU cooler are you using, what is the CPU clock frequency at the beginning of your stress test and during the test, does it change) it sounds like Turbo 2.0 is throttling down during the test, therefore you lose the GFlops you noted and see a slight temperature decrease.
Modern CPUs (and earlier ones as well) are very complex devices and the knowledge base needed to understand their functioning is huge. It is a mistake to immediately assume the CPU is faulty based upon the observations of one type of testing. The Sandy Bridge CPU's are new in the market and users are still learning about their quirks and idiosyncrasy's. One simple thing I can tell you is all socket 1155 mother boards/BIOS are not created equal, and the functioning of the CPU is very dependent upon them, as well as the settings in their UEFI/BIOS.
I've tried everything in the bios settings, i' think.
When all UEFI setting are on default there is also the same performace drop. And it's huge, i think. About -17%. I think it's not a normal operation from any cpu, even if it's throtting. (But there is no reason for throttling or any specific sign of them)
I've some shots from the default UEFI settings and the test:
OC tweaker settings 1 (defaults): http://www.imagebam.com/image/50a2b7131760435
OC tweaker settings 2 (defaults): http://www.imagebam.com/image/fcbdaa131760438
Hw monitor setting (over temp protection on): http://www.imagebam.com/image/b55c4b131760432
Cpu configurations: (default): http://www.imagebam.com/image/2fd961131760443
And the test:
- Nothing interesting, CPU times are OK.
- There is no unnecessary programs running (i've killed all of them before the test)
- Core temps and cpu power are decreaseing under the test.
I've tried non default setting's too:
- Default cpu frequency, but TM1, TM2, overt temp protect disabled.
- Default cpu frequency, C1E, EIST... disabled.
- Default cpu frequency, fixed, increased vcore: 1.25V.
- Manual cpu freq and settings: Turbo short and long TPD limits: 250W, Core current limit:250. Manual multiplier=33.
- Manual cpu freq and settings: Turbo short and long TPD limits: 250W, Core current limit:250, Manual multiplier=44, vCore: Auto (1.38V under linx) / Offset -0.080V (1.30V under linx) / Fix 1.27V (under linx). (with and without other mixed settings, TM1, TM2 disabled...)
There are 3 shots with the last settings, 4.4GHz, TM1, TM2 disabled:
Performance drop: -26%! Max core temps: 74-75C. No freq throttling, no background programs, cpu time is OK, core temp and power falling is the same.
I've searched on other forums, other 2500k's @4.4GHz perform about 108GFlops constantly for hours of linpack running.
Btw: The cooler is an Arctic Cooling Freezer Xtreme rev2.
I' think it would be a temp protect bug. When the performance dropped after 10-15min linx, Cinebench and other tests are performing poor as well. But after 3-5min idle Cinebench and all tests perform normally.
Well it sure seems like a performance drop during the test, without any signs of throttling, which is the weird situation. Going from 103 to 74 GFLOPs is significant and not a minor glitch. I can offer a few comments.
The AIDA64 core temps just prior to the start of the decrease in GFLOPs (the line you have highlighted in LinX) are 69C, 74C, 75C, and 71C (IMO SpeedFan temps are to low.) Tcase for the i5-2500k is 72.6C. You might argue that those temps do not correspond to the Tcase temp, but can we be sure of that with these new CPUs? I've noticed a difference between i7-900 and i7-2600 CPUs in the relationship between their core temps and CPU temp. i7-900 CPUs report core temps about 10C above the CPU temperature, while i7-2000 CPUs do not report core temps above the CPU temp, and are often reported lower than the CPU temp. Anyway, could the temps you are seeing there be the trigger of something? But AIDA64 does not show any throttling.
I've used HWiNFO on Core 2 Duo, Core 2 Quad, i7-900, and i7-2000 CPUs. I think there is a bug in HWiNFO with the Sandy Bridge CPUs, since when monitoring my i7-2600K, the Core Clock bar and frequency reading stays at the max Turbo frequency (matching the multiplier, for example 4200 MHz with a Turbo multiplier of 42), and never changes, while I have EIST enabled, as well as C1E and C3 C-States, and other CPU frequency monitors show my CPU clock going to 1600 MHz when EIST activates at idle. I'm not referring to CPU-Z, whose polling rate seems slow, but the software provided with my mother board (ASUS), and TMonitor.
I've also noticed that during stress testing, hardware monitoring programs seem to lock up at times, not updating/changing. That is not surprising to me, given the CPU is at or near 100% usage during these tests. Yet monitoring like you are doing seems to be a standard practice and I've never read comments about that. My point is, can the monitoring programs be 100% trusted during these tests. Both in their consistent operation and the data they display.
I have also heard of Turbo throttling back during stress testing when there seems to be no reason for it do so. The CPU temperatures were in the 60's C, and the Turbo limiting settings were apparently disabled, and power settings increased, as you have done. But, throttling was seen in these cases.
Sorry to provide at best what are clues and hints at what may be happening. You may need to try another i5-2500K in your system, and see what results you get. If it is the same, and others do not experience the same thing, then it is likely the mother board/UEFI/BIOS is the culprit, of course. I've never experienced what you have with my i7-2600K, but I've never tested it to the extent you have, which I will do now.
I too have your problem with the processor. i5-2500K.
my motherboard is asrock p67 pro3 too.
do you fix a problem ?
How much virtual memory do you have defined? You might try eliminating it for the test (zero VM) so only actual memory is used. Heavy memory use will cause Windows to off-load to VM and you start adding disk I/O to the equation. Could be the drop you're seeing.
Edit: Poor description but Windows needs some of the memory too--For background functions. Also make sure you aren't running un-necessary services.