Iteration times in i5 and i7

[TheMawn]

Originally Posted by Jud McCranie

That should do it. You say that you run the torture test for 72 hours?

I looked up some videos about overclocking i5 and i7, and it may be more involved than I want to get into.

I ran for 72 hours. You get to a point where it's kind of silly to carry on much longer. I bumped myself up a few notches in the voltage department afterward just to be safe. I did a bit of playing around with the settings a few weeks ago and I re-did some torture tests again. 72 hours stable again.

It's never 100% certain but pretty damned close. That's what double-checks are for.

[Jud McCranie]

Originally Posted by TheMawn

... It's never 100% certain but pretty damned close. That's what double-checks are for.

I don't think Seventeen or Bust does double checks (probably does if one reports a prime). Also, I'm going to be doing other things, starting in a few days.

[cuBerBruce]

On my MacBook Pro (Sandy Bridge Core i7), I seem to be seeing Prime95 performance changing significantly from time to time.

For instance earlier this evening, one of the workers was showing iteration times of .040 or .041 fairly consistently for quite awhile. But then, the iteration times suddenly changed to 0.059. I decided to stop the workers temporarily before putting the system to sleep for a few seconds. After taking it out of sleep, I restarted the workers, and remarkably the iteration times were now only .029 or .030, and its still holding that speed over an hour later. By the way, the other workers showed similar changes in performance at the same times. (I have 3 workers, and throttle set to 85%.)

It's frustrating to see Prime95 being slowed down over 2x what it's capable of, and especially with no obvious reason for it.

[kladner]

Originally Posted by cuBerBruce

[SNIP]It's frustrating to see Prime95 being slowed down over 2x what it's capable of, and especially with no obvious reason for it.

Can you monitor temps on the CPU? A rough estimate might be made by feeling how hot the air coming out of it is, but of course, numbers are better. A brief period at idle and in sleep mode could let it cool down and stop thermal throttling, if that's what it was doing when it slowed down.

[cuBerBruce]

Originally Posted by kladner

Can you monitor temps on the CPU? A rough estimate might be made by feeling how hot the air coming out of it is, but of course, numbers are better. A brief period at idle and in sleep mode could let it cool down and stop thermal throttling, if that's what it was doing when it slowed down.

Well, Temperature Monitor reports such wildly varying temperatures for the core temperatures, I pretty much only pay attention to the main heat sink temperatures. I tend to set the throttle duty cycle lower if I see "Main heat sink 3" getting over 50C. It's been in the 48-52 range this evening without making any changes to the throttle setting. A core, on the other hand, can change from say 97 to 73 from one second to the next, for instance, if Temperature Monitor is to be believed. I don't see much point in trying to monitor these wildly fluctuating values.

And the iteration times is often very consistent for long periods of times during the slow times. Would temperature throttling tend to produce inconsistent iteration times?

It's almost seems like the cpu speed is being changed with long intervals of time at a given speed, but the programs I currently have for reporting CPU speed seem to show its speed always at exactly 2.5 GHz (which is also the standard rated speed for the processor). So far I have no confidence in these programs.

[Jud McCranie]

Originally Posted by cuBerBruce

A core, on the other hand, can change from say 97 to 73 from one second to the next, for instance, if Temperature Monitor is to be believed.

I wouldn't believe that is accurate.

[kladner]

Those are some strange circumstances. I don't blame you for being frustrated. If the reported core speed doesn't change, thermal slow down doesn't seem that likely. I have no idea of the response times for thermal control, either. Besides, you're not seeing fluctuations in other parameters to match the rapid changes of temp at the core level.

It might be an interesting experiment to set the on some kind of support to allow more air flow into and around it. If you don't have a notebook cooling pad, even a kitchen wire cooling rack might serve.

Perhaps someone with more Intel experience than I have would have a better idea of what might be happening.

[mdettweiler]

Originally Posted by cuBerBruce

Well, Temperature Monitor reports such wildly varying temperatures for the core temperatures, I pretty much only pay attention to the main heat sink temperatures. I tend to set the throttle duty cycle lower if I see "Main heat sink 3" getting over 50C. It's been in the 48-52 range this evening without making any changes to the throttle setting. A core, on the other hand, can change from say 97 to 73 from one second to the next, for instance, if Temperature Monitor is to be believed. I don't see much point in trying to monitor these wildly fluctuating values.

And the iteration times is often very consistent for long periods of times during the slow times. Would temperature throttling tend to produce inconsistent iteration times?

It's almost seems like the cpu speed is being changed with long intervals of time at a given speed, but the programs I currently have for reporting CPU speed seem to show its speed always at exactly 2.5 GHz (which is also the standard rated speed for the processor). So far I have no confidence in these programs.

I think I've seen the same rapid temperature fluctuations you describe on an i7-2620M laptop of my own. In my case, I was using Core Temp to read temperatures (different programs can give slightly different results but they're usually consistent relative to whatever they think the TJunction Max temperature is, which is how the CPU gives its readouts).

When the CPU isn't under heavy use, the temperature stays steady at an expectedly low level (45 C or so); if I then turn on Prime95 (or another such highly CPU-intensive program) on both cores (it's a hyperthreaded dual-core), the temperatures slowly rise until they reach the 100 C TJunction Max, at which point thermal throttling kicks in and the temperatures start cycling up and down between the mid-80's and the maximum. The interesting thing is that this cycle seems to happen at a much faster pace than on most CPUs - typically, thermal throttling will drop the clock frequency for a few seconds to let the chip cool down, then turn it back up and allow the temperature to slowly rise again to the maximum, at which point the cycle repeats. On other computers that I've observed (both Intel and AMD - the ones I've observed the most are a C2D desktop and a Phenom II X4 laptop), this whole cycle usually takes place over the course of anywhere from 30 seconds to 20 minutes, depending on how well the CPU is doing at venting heat (given factors like dust, etc., which are a big deal especially in a laptop). However, on my i7-2620M laptop, the cycle happens very rapidly - on the order of just a couple seconds between the ~85 C low and the ~99 C high. This confused the heck out of me until I tried reducing Core Temp's update cycle to 100 ms (the lowest) instead of the default 1000 ms; I was getting nonsense readings since the cycle was happening too fast to get useful information at the default interval. At 100 ms, I could see the two cores rapidly (and independently - the two cores' cycles seemed independent as far as I could tell) adjusting their frequencies anywhere in the range of 2.2 GHz to 3.3 GHz or so (which means Turbo Boost is still in play even at these temperatures).

I've noticed that my particular laptop is not the only computer to do this; other i7's that I've tried this on (particularly laptops, since they tend to unavoidably hit their thermal limits at full tilt) exhibit the same throttling behavior at their limits. It appears that the Core i-series processors are able to adjust the clock frequency with extremely low latency, and are very good at dissipating heat quickly at lower frequencies, such that this fast-cycle strategy actually works well. The end result is that the CPU is dynamically adjusting itself to its thermal dissipation capabilities, allowing you to get the most possible performance out of it even under suboptimal thermal conditions.

When I first saw this behavior, I looked up some information online about i7 (or more precisely i-series in general) throttling behavior, and apparently this is to be expected - the i-series processors are extremely resilient to thermal stress, and are designed to safely run all the way up to their 100 C TJunction Max temperature. Unfortunately I do not have the link available as it was a while back and I didn't think to save it, but it did seem to be a relatively credible source - the guy who wrote the article I found worked for some sort of company that specialized in high-performance Intel CPU cooling solutions.

I have never been able to get this laptop to vent well enough for it to not throttle like this under full load (with AVX-optimized Prime95) - it's an ultraportable tablet PC, so I wouldn't expect it to be designed for heavy-duty use like this. If I tilt it on its side to allow better airflow, the fan doesn't have to strain quite as hard as otherwise, and the case seems fairly well-built for thermal stress, so I'm not particularly worried about damaging it over time like this. You may, however, want to be a little careful with your MacBook - other members of the forum have reported that the plastic on their MacBooks has started to warp over time due to long-term heat exposure.

As for the iteration timings, I can't tell you too much since, with the CPU doing this kind of rapid dynamic adjustment to its present thermal abilities, you can't really expect anything to be super-consistent; I've mostly been running work for the Conjectures 'R Us project lately, which has been dealing with a rather wide variety of FFT sizes, so I hadn't been paying much attention to the iteration timings as I'd be comparing apples and oranges. All I can say for sure is that in order to get really consistent timings, you'll need to figure out a way to cool your system efficiently enough that it has a steady temperature (i.e. stays below the 100 C limit and doesn't trigger throttling). If you can do that, you should get full, steady performance out of your CPU. However, it may not be possible since many laptops are not able to achieve this even with a good cooling pad.

One thing you can try, for the sake of experimentation, is to put an ice pack under your computer (yes, really!) - put a thin cloth between the ice pack and the computer to make sure it doesn't sweat and get water into the holes on the bottom. Usually the "gel" type ice packs that can lay flat work best. Unfortunately, this trick doesn't work super well on many modern laptops since they tend to have plastic cases without really good thermal contact between the case and the interior parts, so you just end up chilling the plastic; it might be worth a try though. I have an old Pentium II laptop that this used to work really well on...it had some kind of plastic (might have even been metal) surface on the bottom that conducted heat really well, which was bad when it was on your lap, but made the ice-pack trick work really well. After 5 minutes of being on the pack the whole thing would be cool to the touch with the fan running at lowest speed. Long story short - your mileage may vary.

[cuBerBruce]

Thanks (especially to mdettweiler) for the comments. mdettweiler's explanation seems to correspond with how my system appears to be behaving. I had had the impression that thermal throttling generally caused throughput to get reduced drastically. But, at least with some Core i7s, as mdettweiler describes, apparently thermal throttling does not necessarily result in bad performance. I note that even though my MacBook Pro appears to throttle like this while running Prime95, it still outperforms (by a big margin) any of my other computers running Prime95. The processor is a Core i7-2860QM.

Originally Posted by mdettweiler

I have never been able to get this laptop to vent well enough for it to not throttle like this under full load (with AVX-optimized Prime95)

I don't seem to need to be running Prime95 anywhere near "full load" to get this behavior. I doubt it's possible to run it at any "useful" speed and not see this.

When I started running Prime95 on this laptop, I was concerned with seeing temperatures in the 90s a lot. I decided to run only two workers and sometimes turned one (or both) off when the heat sink temps were getting higher than I preferred. A few weeks ago I added a third worker, but used the throttle setting to limit workload rather than turning workers off. In the near future, I'll probably try going to a 4th worker (and decreased throttle setting).

Originally Posted by mdettweiler

You may, however, want to be a little careful with your MacBook - other members of the forum have reported that the plastic on their MacBooks has started to warp over time due to long-term heat exposure.

Apple says the case is made of aluminum. And yes, it has gotten hot enough that I have to move it away from my skin. I generally have the back of it hanging over the edge of a table/chair so it can vent better. Doing this seems to help bring down the temperature. I imagine it needs some cleaning. I've been reluctant to try to remove the cover.

- - - - - -

I note that just before going to bed last night, the iteration times of the workers jumped up again. I did the sleep trick, and the iteration times immediately dropped down again. They stayed down over 4 hours, then jumped up again. I tried the sleep trick again, and the iteration times are down again. So anyway, it seems like I have a way of dealing with the slowdown once I notice that it has occurred. I'll try cleaning it and some other experiments when I have time.

I'm guessing after throttling like it does for awhile, it decides to limit the max frequency it will run at or something like that, causing a noticeable change in performance level. Perhaps putting it to sleep resets some state machine controlling this, so it goes back to the default max frequency. I'm just guessing, though. Anybody know if there is a way to get the true running frequencies of the cpu (or individual cores, if they can run at independent speeds)? (Remember this is Mac OS X - Lion, not Windows or Linux).

[mdettweiler]

Originally Posted by cuBerBruce

Apple says the case is made of aluminum. And yes, it has gotten hot enough that I have to move it away from my skin. I generally have the back of it hanging over the edge of a table/chair so it can vent better. Doing this seems to help bring down the temperature. I imagine it needs some cleaning. I've been reluctant to try to remove the cover.

You actually shouldn't need to remove the cover to properly dust out a laptop - they have their airflow optimized such that there's essentially one unified airflow path through the case. It's not like a desktop where the dust can get trapped in all manner of nooks and crannies. Blowing compressed air in through the keyboard should cause the dust to come out via the fan vents and intakes. (Beware - this will give you a big nuclear cloud of dust, so do it somewhere that it won't get all over things. And maybe put something over your nose and mouth. )

I note that just before going to bed last night, the iteration times of the workers jumped up again. I did the sleep trick, and the iteration times immediately dropped down again. They stayed down over 4 hours, then jumped up again. I tried the sleep trick again, and the iteration times are down again. So anyway, it seems like I have a way of dealing with the slowdown once I notice that it has occurred. I'll try cleaning it and some other experiments when I have time.

I'm guessing after throttling like it does for awhile, it decides to limit the max frequency it will run at or something like that, causing a noticeable change in performance level. Perhaps putting it to sleep resets some state machine controlling this, so it goes back to the default max frequency. I'm just guessing, though. Anybody know if there is a way to get the true running frequencies of the cpu (or individual cores, if they can run at independent speeds)? (Remember this is Mac OS X - Lion, not Windows or Linux).

It might not have anything to do with any states being reset (at least not directly); putting the computer to sleep gives the CPU a "power nap" of sorts. Since the CPU is completely turned off (not just idling) it can cool down rapidly, so that when it comes back up the throttling is no longer an issue. Of course, it won't be long before it's throttling once more; but it does give it a little break. This technique can be useful when a laptop is in imminent danger of overheating and is unable to "save" itself quickly enough even under idle due to built-up dust and other factors. (I've had to use it a few times on my AMD Phenom II X4 laptop after subjecting it to intense CPU+GPU usage, i.e. gaming - even when you shut off the game and leave it at idle, the GPU is still putting out too much heat for the CPU to keep up with its own heat output even at its lowest throttling frequency.)

What's interesting about your situation is that you're getting 4+ hours of increased performance after the sleep-trick "power nap". Without knowing the actual temperature readouts, I can only speculate as to why this is happening, but you may be on to something with your theory about the CPU deciding to implement a temporary frequency ceiling after running at full load for an extended period. My Phenom laptop does something similar if it gets too dusty - when you first boot it up, it'll try cycling all the way up to the full clock frequency (2.2 GHz), but when it realizes that it can never hold that frequency for any length of time it just gives up and stays at 1.6 GHz indefinitely.

As for reading the real-time frequencies from the CPU, I know most popular temperature monitoring programs on Windows can do this (Core Temp, RealTemp); but I'm not familiar with what's available on Mac. On Linux, the usual solution is to use the lm-sensors package, which provides an interface for a wide variety of readouts; you can run "sensors" from the terminal to get a basic readout, or use a variety of GUI frontend programs.

[Jud McCranie]

Originally Posted by mdettweiler

What's interesting about your situation is that you're getting 4+ hours of increased performance after the sleep-trick "power nap".

Yes, but will the increased performance make up for having it off for a while?