Iteration Times Vary on Core i7

[AZMango]

Originally Posted by axn

Did all 4 LL tests show .035 simultaneously?! (Doubtful). Your timing of 0.055 is more like what is expected of your CPU. Hyperthreading is not at all useful for P95. Running 4 single-threaded LL is the most efficient use of your CPU (as far as P95 is concerned).

One test is working on a 3072 FFT length so that one was a bit higher than the other three. Right now I'm getting about .055, .055, .067, and .055 on my four tests.

I wonder if maybe those .035 times weren't accurate for some reason since my expected completion dates did jump ahead about six days, but was slowly going back to where it was the longer it was running that way. Would be nice if those lower times showed up again so I could run the coretemp program to see what was going on.

I thought I saw some benchmarks that showed Core i7 920's at about .040 or less, but maybe those were overclocked. Mine isn't.

[AZMango]

Originally Posted by petrw1

Did you say the 6 degree temp drop made no difference on the iteration times? Seems odd? 72-76 is not bad but a little higher than I would like.

If you routinely see Prime95 at about 100% (or 50% without Helpers) then that rules out other things stealing cycles.

Bigger coolers are not necessarily louder. The one recommended to me (Noctua) for my i5-750 is quieter than the standard cooler/fan on my other quad.

Not sure how you define "big cooler" but mine was big enough that I could not close the case...though bigger does NOT necessarily mean better cooling either.

But yes I would say that according to others benchmarks with the i7-920 you should be around 0.035 with 4 LL in the 30-35M range.

http://www.mersenneforum.org/showthr...d=1#post208318

By big I just mean that it seems to be quite large, with a big tube on it guiding the air in. It's extremely quiet, so it's probably not putting a whole lot of air through.

Is that .035 range without overclocking? I read that the i7's have Turboboost mode so found a gadget that shows the current speed and it's at 2.79, same thing Core Temp shows. Maybe one of these days when I reboot again the faster times will show up, I'll check things out then.

Not sure I want to mess with the cooler at the moment since I wonder if it might void my warranty for some reason. But since the times don't seem to change with a lower temp, not sure it would matter.

Thanks for the help.

[petrw1]

Originally Posted by petrw1

But yes I would say that according to others benchmarks with the i7-920 you should be around 0.035 with 4 LL in the 30-35M range.

http://www.mersenneforum.org/showthr...d=1#post208318

Hmm...cut and pasted the wrong URL. This is the benchmark one I was talking about:

http://www.mersenne.org/report_bench...Get+Benchmarks

You'll note that the other i7-920 (not OC) has iterations times of 51 for 2560 and 62 for 3072K FFT.
I had assumed you were running a smaller FFT when I made my above comment. Sounds like you are in the ballpark. Crunch away!!!

[AZMango]

Originally Posted by petrw1

Hmm...cut and pasted the wrong URL. This is the benchmark one I was talking about:

http://www.mersenne.org/report_bench...Get+Benchmarks

You'll note that the other i7-920 (not OC) has iterations times of 51 for 2560 and 62 for 3072K FFT.
I had assumed you were running a smaller FFT when I made my above comment. Sounds like you are in the ballpark. Crunch away!!!

Thanks for the help. I guess .055 isn't too far from the .051 they show. If I see those .035 timings again I'll let you know.

[cheesehead]

Originally Posted by AZMango

So, are you saying that even though it says 52%, it's really 100%? How about when the task manager thinks it's 100%?

For the purpose of illustrating my statement about load percentages, but not to be a technically precise description of how hyperthreading works:

Task Manager (a) "thinks" there are two CPUs, each with the same maximum load capacity, and (b) displays load percentages as though each CPU can have only 50% of the total at most.

The (only, in reality) real CPU can be treated as though it were a different ("virtual") CPU while it is executing hyperthreaded instructions than it is while executing nonhyperthreaded instructions. So whatever goes on when the nonhyperthreaded thread is executing is counted toward the real CPU's total, and whatever goes on when the hyperthreaded thread is executing is counted toward the virtual CPU's total. It's an accounting scheme to allocate "load" to the real CPU and virtual CPU separately.

When the real CPU is fully loaded and the virtual CPU is "idle", that is displayed as a 50% ((100 + 0)/2 = 50) total load by Task Manager.

52% total load could mean that the real CPU is 96% loaded and the virtual CPU is 8% loaded ((96 + 8)/2 = 52), or the real CPU is 100% loaded and the virtual CPU is 4% loaded ((100 + 4)/2 = 52), or that both the real and virtual CPUs are each (independently) 52% loaded ((52 + 52)/2 = 52).

100% total load means that both the real CPU and the virtual CPU are fully loaded ((100 + 100)/2 = 100).

[Prime95]

Originally Posted by AZMango

If I see those .035 timings again I'll let you know.

The .035 times most likely came from prime95 briefly (and incorrectly) running a worker on two physical CPUs rather than the two hyperthreads of one physical CPU.

[AZMango]

Originally Posted by cheesehead

When the real CPU is fully loaded and the virtual CPU is "idle", that is displayed as a 50% ((100 + 0)/2 = 50) total load by Task Manager.

52% total load could mean that the real CPU is 96% loaded and the virtual CPU is 8% loaded ((96 + 8)/2 = 52), or the real CPU is 100% loaded and the virtual CPU is 4% loaded ((100 + 4)/2 = 52), or that both the real and virtual CPUs are each (independently) 52% loaded ((52 + 52)/2 = 52).

100% total load means that both the real CPU and the virtual CPU are fully loaded ((100 + 100)/2 = 100).

This kind of makes sense, but is still a bit confusing in regards to Prime95. I think I understand why it is that with two CPU's if one is 100% and the other is 0, then the average is 50%. Checking Task Manager with Prime95 using only one CPU each my percentages vary with each pair acting as opposites. If the real CPU is at 80%, then the virtual CPU is at 20%. One is about 40%/60%. Of course there are corresponding spikes in each pair. Kind of like mirror images if they were one above the other.

One thing I wonder about is that since I have the setting to use just one CPU, it still says, "Setting affinity to run worker long logical CPU's 4,5", for example. Does this mean my PC is only being 50% used? Does setting the multithreading CPU's to '2' really cause things to be 100% overall? Or, is nothing really being gained by having the virtual CPU's help? At least pertaining to Prime95 since it's already using both CPU's of each core? Maybe if I sleep on it it'll make more sense. Thanks for the help.

[AZMango]

Originally Posted by Prime95

The .035 times most likely came from prime95 briefly (and incorrectly) running a worker on two physical CPUs rather than the two hyperthreads of one physical CPU.

Would that still be the case with four LL's being done at the same time? All four of them were showing under .040.

[lfm]

Originally Posted by AZMango

By big I just mean that it seems to be quite large, with a big tube on it guiding the air in. It's extremely quiet, so it's probably not putting a whole lot of air through.

That sounds like the special Dell cases. I think Dell has their own heat sink design to fit it too. Custom heat sinks might be trouble to fit and probably will require you remove the air tunnel/tube which could actually wind up worse overall in the end.

I wouldn't be surprised if Dell would object to replacing that heat sink whether its justified or not. They have engineered the case, heat sink and the air tunnel to work together.

On the other hand Dell might be trying to use an old case and heat sink design that is marginal for a new processor like yours.

[cheesehead]

Originally Posted by AZMango

Checking Task Manager with Prime95 using only one CPU each my percentages vary with each pair acting as opposites. If the real CPU is at 80%, then the virtual CPU is at 20%. One is about 40%/60%. Of course there are corresponding spikes in each pair. Kind of like mirror images if they were one above the other.

My mistake -- what I get for not doublechecking before posting.

[joblack]

Originally Posted by Prime95

The .035 times most likely came from prime95 briefly (and incorrectly) running a worker on two physical CPUs rather than the two hyperthreads of one physical CPU.

Does Prime95 recognize HT cores? It might be that the process scheduler put the prime threads on the wrong (virtual) cpus. Might be counterproductive if it puts two threads on one physical core.