Originally Posted by Faithh
How can you notice a performance decrease while you're being focused killing people in BF3? A few runs of 3dmark 11 isn't going to dig that out. You're never ever going to notice there happened a computational error or inaccurate calculations by your CPU. Because you have an error doesn't even mean it's going to crash, hence the message from prime "Worker stopped took too long" whatever it was.
This might be interesting to read for you, even a bug in a CPU won't always cause crashes: http://en.wikipedia.org/wiki/Pentium_FDIV_bug
I talked with Forceman over in my Haswell Overclocking Thread. I still think I'm right. Only Forceman has talked about this issue, others never even heard of this. So I'm going to copy paste the exchange so far:
"It can, if the chip is doing error correction. If it recognizes the error and then re-does the calculation to get it right, it'll affect performance. You can sometimes see it in the GFlops on IBT.
Interestingly enough, I've found (in the many hundreds of times I've run IBT) that there is definite voltage scaling to the performance, at least on my chip. At too low a voltage it crashes, but give it just enough voltage to pass and you get 125 GFlops (for example). Increase the voltage a little more, and it'll go up to 126 or 127 GFlops. Increase the voltage even further, and it'll start going back down a little, like to 126.5 or 125.5. The increase from the initial voltage bump makes sense from an error correction standpoint, but I don't get why it goes back down again."
And that is the performance difference of what, 0.8%? even if that is the case. No FPS drop will be noticed. The person said we won't notice the FPS boost of further overclocks (extra 100mhz after wall where you can't test with synthetics), yet we'll notice the performance change of less than a percent? Argument makes no sense.
And in addition to all that, if one cares enough for that 0.8% the person will test the performance over and over to make sure the CPU is performing up to snuff so again the argument is moot.
Hard to imagine as well, CPU never, ever Bsoding when it's exhibiting such errors. Say, BF3. Or Stockfish 4. Ough to trigger Bsod one of those days.
@darkwizzie - you don't see it as much with Haswell for some reason, but with Ivy you used to see people getting lots of non-crashing WHEA errors when testing, which is basically a successful error correct by the chip. Bumping the voltage would get rid of them, so it is definitely stability related. I agree it's a pretty marginal performance difference , but like you pointed out, eventually one of those errors will be uncorrectable and then you'll get a 124 error.
Therefore I still think I'm correct.
If the issue is FPS dropping large enough to be tested and noticed in an average video game, video game being more GPU reliant as a rule, that highlights a HUGE decrease in CPU performance, very easily detected by a chess benchmark. If the FPS is too small to be measured reliably then the difference is so small it doesn't matter.
If you crash then obviously it's not stable and things need to be changed.
For your argument to work, the change needs to be undetectable. That means not crashing, restarting, bsoding, lockups, etc, ever.
With a chess benchmark, it was sensitive enough to catch changes in uncore. That's pretty damned precise. If I can't measure a decrease in performance when I went 4.5 to 4.6 then absolutely the difference is too small to matter.
You said it won't cause crashes at the start IIRC, too lazy to check. Then you say "won't always cause crashes" well doesn't that render the whole point of your original argument pointless if it sometimes crashes? And first you said performance will cause FPS drops, meaning on an average game that means a big change in CPU power, which would be sufficient to cause measurable performance changes and therefore can be detected and stopped. Then you say it might not be detectable even in CPU benchmarks. Then isn't the performance change so small, the argument is arguing over something that won't make a practical difference?
Edited by Darkwizzie - 9/16/13 at 11:37pm