Real Temp - New temp program for Intel Core processors

[unclewebb]

http://www.fileden.com/files/2008/3/...alTempBeta.zip

I've added the ability to run a file when the RealTemp alarm is triggered.

I haven't decided yet if I'm going to add a file picker into the Settings GUI or keep it hidden. At the moment you can try this feature out by adding a new line to the RealTemp.INI configuration settings file like this:

AlarmFile="C:\Program Files\IrfanView\i_view32.exe"

You can probably think of something better than running IrfanView when you're computer is getting ready to go nuclear!

The AlarmFile will only be run once when an alarm occurs. If you want RealTemp to run it again then you'll have to restart RealTemp. A person could also write a .bat file for an orderly shut down or whatever they want their computer to do during a high temp alarm. Any suggestions about improvements or additional options for this feature are appreciated.

I'm in the middle of adjusting the Start Minimized feature in RealTemp for better Vista x64 compatibility. If that part of RealTemp 2.78 is working worse than before in Vista then don't be surprised. It's a work in progress. If I can't make it better then I'll at least go back to the old way. Version 2.78 is for users that want the AlarmFile feature right NOW!

genec57: I know a little bit about sticking core digital thermal sensors but my experience with a sticking CPU sensor is the same as yours. Zilch!

With any sensor, if you know it is sticking, even sometimes, then you can't rely on it to give you accurate and usable information below that sticking point. The Intel on chip sensors are consistent and usually fine when the temperature goes above the "bottoming out" point.

Same goes for you radaja. The RealTemp calibration feature is not designed to give you usable information out of a stuck sensor. If the sensor isn't moving at low temps then low temperature information coming from this sensor is useless.

I might be able to understand your situation better if you send me some data. Turn on the RealTemp logging feature and set the interval to 1 second. Let your computer sit idle for a minute, then run Prime95 small FFTs for about 2 minutes and then stop that and return your computer to idle for a minute. Take the log file and either e-mail it to me at the address in the About... box of RealTemp or copy and PM it to me here at XS. I can't guarantee you any miracles but I've seen enough messed up data that I might be able to help.

[rge]

Shutdown feature works great and definitely useful addition

I have vasilios shutdownXP enforcer, free program that will do a hard (immediate) shutdown.

If anyone wants to use it, search bolded name, available at many sites. Install it (nothing runs in background, it only runs if activated to shut off pc). Add one line of code to realtemp ini file, exactly as below in bold (unless default install location has changed)

AlarmFile="C:\Program Files\Vasilios Applications\ShutdownXP Enforcer\Data\Shut.exe"

Set alarm temp in settings (I set mine to normal load temp to test). Works perfect!

[unclewebb]

Nice find rge.

I'll include a link in the RealTemp docs to Vasilios ShutdownXP Enforcer v6.3

The start minimized feature is presently more broken than usual in Vista. Now that the alarm feature is more useful, I can get back to working on the start minimized issue that still effects many Vista users.

[AndrewZorn]

I took my Q6600 on water down to 1600mhz and 1.2v (couldn't figure out the 1.08v EIST thing...). I wasn't sure of the adjusted value. The read says 5*C for a dual, 6-7 for a quad. But I'm on water, but also my voltage is higher. I added 5*C to room temp. Is this right?

My adjustments were -4, -2, 2, -2. This made it 27*C across the board at ultraidle. Core0 and Core2 used to be 10*C apart, all the time, idle or lots of load. This doesn't really make sense. Now all my temps are within 4*C of each other during load. Is this more realistic, or am I fooling myself? I don't see how no matter the load two near-adjacent points are 10*C different, so this makes sense to me.

Overall load temps seem lower when compared to Core Temp, that's why I'm not sure if my +5*C over room temperature is right.

Great program either way!

EDIT first two cores are still hotter during load, which leads me to believe it's possible that my sensors were poor just like all of them BUT that I still have a realistic difference between the quality of the two dies.
EDIT and still a ~7*C max difference between different cores, which I can believe... but can't necessarily trust. Waiting for someone to call me out!
EDIT also forgot to say, the "sensor test" also waaaaay evened out, a perfect 3-3-3-3 at ultraidle, from something more like 1-7-4-you-get-it.
EDIT and now things make more sense, as Core2 used to be the coolest of cores but would also be one of the first to fail every time. Now it is one of the warmest cores...

[radaja]

@unclewebb thanks for the E-mail, i was running 450x9@4050 with 1.38v on my E8500.

[unclewebb]

AndrewZorn: I found this test kind of interesting on my Q6600 - G0:



Run Prime95 small FFTs and for "Number or torture test threads to run:" enter 3 instead of 4.

When your CPU is only at 75% load, it constantly moves that load around and it seems to be based on the data coming from the digital thermal sensors. I think when a process asks to run on any core then the CPU runs it on the core that is reporting that it is running the coolest. There are two Dual Cores within a Quad so I think the load is first split in half (more or less) and then those two loads are shared by core0 / core1 and by core2 / core3.

When uncorrected, my core2 always reports the highest Distance to TjMax value which means it is running the coolest so in this test it gets the biggest load and its partner, core3, gets to take it easy. The sensors on core0 / core1 always report very similar values so their half of the load is shared much more equally between those two cores. In both situations, the Distance to TjMax is being balanced for both sets of cores. In theory, Core i7 should be able to move the work load around and balance the temps between all 4 cores at the same time but the design of the Core 2 Duo seems to prevent this.

Task sharing by the CPU seems to be directly based on data from these temperature sensors. When you have a core reporting that it is running cool, it is going to get a bigger share of the work load. If in fact it tends to run at the same temperature as its partner but is misreporting its status, then it's going to be forced to run a bigger load which will make it run a little hotter than its partner. You can tell by looking at the graph that this balancing is going on constantly. Sensors that are not well balanced from the factory can lead to individual cores running at slightly different temperatures.

At higher temperatures where these sensors are much more accurate, the effect of the above declines and you will typically see core0 / core1 running at the exact same temperature and core2 / core3 also running at the exact same temperature. In most of the Q6600 screen shots I've seen, it's almost always core0 / core1 that report that they are running hotter. Typically about 0C to 5C which is shown in the above screen shot.

This might be just a sensor issue but it doesn't seem to be. Think about what would happen if during manufacturing that there was a slight difference between how the IHS contacted each set of cores within a Quad. It's possible that two of the cores might be making better contact with the IHS than the other two so the heat transfer is more efficient and they end up running cooler. Go try my test on your Quad and see what you get.

radaja: Your temps are great when running at that speed.

[wongnog]

Have any of you guys tried out Intel Burn Test from this forum? It gets my Q6600 G0 super hot, I can easily get within a stone's throw from Tmax (100°C) depending on my OC and voltage. I wonder if my cooler (OCZ Vendetta 2) is not mounted well... Anyway I dialed down my voltages and OC so my core temps only reach ~82°C max just to be safe.

[unclewebb]

Intel Burn Test definitely creates some big heat, especially on a Quad. It's great to be Burn Test stable but if you don't ever use any software that ever gets your processor this hot then maybe it's a little overkill. I thought Prime small FFTs was a little overkill on most Quads compared to the temperature that most software runs at but Intel Burn Test is WAY overkill. I still like it and use it though!

[Praz]

AndrewZorn: I found this test kind of interesting on my Q6600 - G0:

Those results may be a bug in Prime95. It did the same thing when one core would fail. After I brought this up the issue was fixed in v25.6. Might want to ask over at the Mersenne.org forum for confirmation one way or the other.

[AndrewZorn]

UncleWebb, that is actually very informative. I will try it when I get home!

However, I wasn't referring to the different dies being of different temperatures, as I see that even after the correction. What I mean is there is sometimes a 10*C difference between individual cores on the same die, even at super-duper-idle-near-room-temp or overkill load. After adjustment the values seem to "make sense" and I'm getting more of what you were explaining in the post.

I also wonder if +5*C over room temp is good for a Q6600 at 1.2v 1600mhz on good watercooling.

[unclewebb]

Those results may be a bug in Prime95. It did the same thing when one core would fail. After I brought this up the issue was fixed in v25.6.

I was using version 25.6 in that screen shot. It's not a bug in Prime, just the way Quad core and Dual Core processors balance the load. The same thing happens when you have a Dual Core and run 1 instance of Prime. The CPU will constantly swap the task back and forth between cores to keep the cores balanced and core temperature seems to play a big part in that decision. This also happens when you run the typical single threaded game on a Quad. It doesn't run one core at 100%. It constantly spreads the work around so each core does part of the job so no one core gets over stressed.

AndrewZorn: JohnZS may not have the record but his 45nm Quad certainly ranks highly when it comes to unbalanced sensors. On the same die he had a difference of 7 between core2 and core3 and the difference was 15 between the two dies when comparing core1 and core2. When you pull out the ruler and measure the number of millimeters between sensors, it becomes obvious that something isn't quite right.

Intel showed us that their 45nm sensors suffer from slope error where the further you get from their calibration point, the bigger the error can become. If you have one sensor that reads too high and one that reads too low on the same die then the combined error can be quite significant.



Intel did not get specific about what temperature they calibrate at but the example in the graph shows how slope error grows as temps decrease.

In my experience, most of the 65nm CPUs I've seen have had sensors on the same die that were either both too high or both too low. The 45nm Quad processors, especially some Extreme CPUs, can have both high and low reading sensors on the same die.

65nm sensors can also bottom out or get stuck at lower temperatures but this seems to be much rarer and at far cooler temperatures than some of the 45nm sensors get stuck at.

If you are water cooled then you need to be comparing your Ultra-Idle reported temperatures to your water temperature. (I like the term Ultra-Idle for low MHz / low volts.) Is there any way for you to manually set your core voltage lower in the bios? Try that and enable C1E or lock your multi to 6.0 when testing if your computer doesn't complain. My Asus boots up fine but some boards don't like to boot up when locked to 6.0x

I think a Q6600 being 5C or 6C above your water temp during this test is pretty reasonable. Open your case up and send me a log file at a 1 second interval of running 1 minute idle, 2 minutes of Prime small FFTs on all cores followed by 1 minute of idle. That helps me understand your processor better than just seeing a single screen shot.

[Praz]

I was using version 25.6 in that screen shot. It's not a bug in Prime, just the way Quad core and Dual Core processors balance the load.

You may very well be correct. However, there should be no difference between selecting three cores or Prime running on three cores because the fourth failed. With v25.6 when one core fails the load drops to zero on the failed core and the remaining three stay at 100%. Previous to v25.6 a failed core resulted in the same behavior you are seeing with selected cores.

[jason4207]

Actually the low watts and the below intel doc may be close to making sense with the 5C gradient we measure...if the chart below from intel is correct, which implies the gradient from core to casing increases (doubles) when removing the heatsink.
http://www.intel.com/support/process.../CS-011039.htm

Tjmax-tcasemax = theta (jc) x TDP
Solving for theta with Tjmax 100 = .43 C/W

If the thermal resistance from core to casing (theta) does increase/double without a heatsink as intel states, then if an actual 1.6-2.5C gradient does exist we would measure it as 3.2 to 5C with heatsink removed.

Or, stated another way, instead of actual gradient (heatsink on) = .43 X TDP, gradient we measure with heatsink off = .86 X TDP. At idle, extended halt intel lists 8W, deep sleep 6w, and at idle undervolted, underclocked probably more in 6W or less range...clearly this will be variable cpu to cpu and some are testing even lower. But at 6W (max?), the gradient would be 2.5C, we would measure 5C, assuming the intel doc is correct. If true, that would suggest tjmax is 100, when we measure the IHS with heatsink off we would measure 95C, if we could measure the IHS with heatsink on we would measure 97.5, which would make sense with the actual ~2C gradient. But other than a calibrated thermocouple embedded in IHS...dont know how to test that.

Uncleweb, do you have any plans to try embedding a thermo coupler into the IHS (or maybe a HS base) based on this info?

[unclewebb]

Praz: I understand what you're saying. It looks like in Prime when one core fails then 3 cores run at 100% and 1 at 0% but if you tell it to run 3 instances of Prime on a Quad then you get the four cores working together on the problem.

do you have any plans to try embedding a thermo coupler into the IHS

I can't say it hasn't crossed my mind.

I'm not sure what new information we'd end up with though. We know TjMax for 45nm and based on that and how 45nm IHS readings compare to 65nm, I think we can be pretty confident with TjMax for the majority of the 65nm CPUs so what's left? Intel has told us about slope error so we know that variable exists and we have a way to correct for that.

About the only thing that I don't have any hard evidence to prove is how many degrees hotter should a core be reporting compared to the ambient temperature or water temperature during the Ultra-Idle test. Given that a good water set-up or a good air cooler or even the Intel OEM cooler with the copper core all perform within a couple of degrees of each other during Ultra-Idle, I tend to think that about 5C to 7C over ones baseline air/water temperature is pretty reasonable. Maybe an embedded sensor might get us another degree closer to the truth but given the less than perfect nature of these sensors, I'm having a tough time seeing the point vs cost of that.

With there being some unknown error in how Intel sets TjMax as well as slope error, an individual test on 1 or 5 or 10 processors won't be enough data to confirm anything.

With these sensors, the reported temps are about as real as they're ever going to get. Intel didn't release enough information at the IDF, and I don't imagine they ever will, to get us any closer to the real temperature than what we are now.

[jason4207]

I can't say it hasn't crossed my mind.

I'm not sure what new information we'd end up with though. We know TjMax for 45nm and based on that and how 45nm IHS readings compare to 65nm, I think we can be pretty confident with TjMax for the majority of the 65nm CPUs so what's left? Intel has told us about slope error so we know that variable exists and we have a way to correct for that.

About the only thing that I don't have any hard evidence to prove is how many degrees hotter should a core be reporting compared to the ambient temperature or water temperature during the Ultra-Idle test. Given that a good water set-up or a good air cooler or even the Intel OEM cooler with the copper core all perform within a couple of degrees of each other during Ultra-Idle, I tend to think that about 5C to 7C over ones baseline air/water temperature is pretty reasonable. Maybe an embedded sensor might get us another degree closer to the truth but given the less than perfect nature of these sensors, I'm having a tough time seeing the point vs cost of that.

With there being some unknown error in how Intel sets TjMax as well as slope error, an individual test on 1 or 5 or 10 processors won't be enough data to confirm anything.

With these sensors, the reported temps are about as real as they're ever going to get. Intel didn't release enough information at the IDF, and I don't imagine they ever will, to get us any closer to the real temperature than what we are now.

Yeah, I definitely have the same thoughts as you. Right now the only benefit to the embedded coupler test is to perhaps come up w/ a better ultra-idle (I like the name too!) calibration number. I know it's near impossible to say definitely what that number is in every case, but for some reason I've been obsessing over it. I also think it would be nice to corroborate what rge posted.

I think we could make a somewhat meaningful conclusion w/ only 1 CPU, though. If you measure a given CPU w/ an embedded TC w/o a HS, and then add the HS you could see the difference and then perhaps come up w/ a slightly more accurate ultra-idle calibration number.

Also, I wonder if it would skew the results if the TC was embedded in the HS instead. That way you would only need to destroy 1 HS, and use it on multiple CPU's.

Please don't feel obligated in any way to do any of this. It would be interesting, but I'll still live a long and fruitful life w/o this info.

Thanks again!

Jason

[unclewebb]

I think Intel's lack of disclosure about how much TjMax varies from one processor to the next has tied our hands. If TjMax varies by plus or minus one or two degrees due to factory calibration inaccuracies, etc. then there's no way we can come up with a perfect number to shoot for when we're trying to calibrate RealTemp to an individual processor. The recent Intel info posted by rge finally puts all the pieces together and helps make sense out of the data we've both gathered.

Six months ago I had a handful of processors and with the software, Intel documentation and knowledge available, there was no easy way to compare them all. I made some assumptions, some right, others not so right, but I think what we've all learned since then has us much closer to being able to make meaningful temperature comparisons between 45nm and 65nm, Dual or Quad Core.

Carving up an IHS or heatsink still sounds like fun though. I'm sure one day rge or I or some other XS type will fire up the Dremel and have at it! I might have to go price out some thermocouples so we can all have a good night's sleep.

[AndrewZorn]

UncleWebb,

Thanks for the info. I need to set temp to 5-6*C over WATER temp? That might be higher than I hoped. I have a temp sensor thing with a probe on a wire, so I'll shove that down my T-line.

Default voltage is the lowest I can go on this board, at least in this BIOS. I need to see if I can find C1E (that's what it is... I tried something else).

[Dee]

UncleWebb,

Thanks for the info. I need to set temp to 5-6*C over WATER temp? That might be higher than I hoped. I have a temp sensor thing with a probe on a wire, so I'll shove that down my T-line.

Default voltage is the lowest I can go on this board, at least in this BIOS. I need to see if I can find C1E (that's what it is... I tried something else).

AZ!

Hello. Ninja here.

[WoodButcher]

UncleWebb,

Thanks for the info. I need to set temp to 5-6*C over WATER temp? That might be higher than I hoped. I have a temp sensor thing with a probe on a wire, so I'll shove that down my T-line.

Default voltage is the lowest I can go on this board, at least in this BIOS. I need to see if I can find C1E (that's what it is... I tried something else).

If you have the vid card in your sig on that loop you'll need to add much more than 5C to your numbers

[rge]

I bought a Fluke with 2 thermocouples several months ago when I bought E7200, with intention of putting one in IHS and after measurements vs software, then second on core to measure versus IHS...but after learning about a few issues...and killing my E7200 within minutes of initial temp testing...

Thermocouples are very inaccurate for surface temp measurements, unless of course the surface temp=air temp, because if you lay one on surface 3/4 is sensing air temp, 1/4 sensing surface. When my IHS reads 95C by IR, the thermocouple will bounce around from 60 to 70C, but immerse the tip completely in boiling water and it is very accurate. Now I know why intel fully embeds it in the IHS, have to for accurate IHS temp. Even after you imbed it in IHS, it is now sensing 2/3 IHS and 1/3 thermal (?) adhesive (?) you used to seal it, and you have to recalibrate it by accurately measuring the surface temp of IHS at same time, or it will be off a few C (that is apparently why they use second source measurement of IHS in all research papers). And since it can not measure surface temps, I would have to embed my second one in the core or solder it to the side and then calibrate thermocouple to surface temp core. I dont know if that would kill cpu or not (had I not killed it temp testing, I was going to give it a go as the last test). Reason for having to embed the second one in the core, would be because I dont know down to 1C accuracy what the core temp is, hence I cant know the gradient. And since it is near impossible to test temp of core, because once you remove the IHS the temp rises so rapidly, you end up "guessing" at temp, definitely not accurate enough for 1C calibration.

Given intels recent chart...I could now compare IHS temp to thermocouple temp with heatsink off (versus software reading core temp as control), then use thermocouple to measure gradient with heatsink on (using software core as control) and then would know the difference in gradient with heatsink on or off, but only within 1C of error or more if calibration is different for the thermal adhesive I use with heatsink cooling it....but again that is not going to give me actual core temp, because I dont know exact temp of core at tjmax.

Also if I embed one in the core (without killing it) or soldered to side + surface core recalibration, and embed one in IHS, not sure if it will change temp 1-2C after my piss poor die reattach substitute. Thought about drilling hole, but then how do I solder to core. And to give an idea of how hard it would be to calibrate the thermocouple on core after soldering it to the side (since 1/2 thermocouple is reading solder temp)... when I delided the E7200, I turned computer on with delided IHS pressed up against the core, immediately removed once computer running and temps went from 60 to 138C before smoking and dieing (by realtemp core reading) in several seconds despite powerful box fan blowing, though quickly looking back and forth between IR gun and realtemp and getting two pics flashed, i was approximating a tjmax of 90C. Since I know tjmax is higher now, either realtemp and IR never had chance to read same temp at same time or I changed the resistance so much by removing the obviously very effective core cooling IHS, that cental core was hotter than its surface, like at load.

I was only half kidding in an earlier post about an MRI temp...I know you can read tissue temp to within 1C, without disturbing anything..that would be the way to accurately measure temp while determining exactly what you were measuring at same time, without destroying die attach.

[jason4207]

UncleWebb,

Thanks for the info. I need to set temp to 5-6*C over WATER temp? That might be higher than I hoped. I have a temp sensor thing with a probe on a wire, so I'll shove that down my T-line.

Default voltage is the lowest I can go on this board, at least in this BIOS. I need to see if I can find C1E (that's what it is... I tried something else).

Try setting the voltage to auto, and leave your multi on auto. Also make sure anything labeled speedstep, EIST, or C1E is enabled. Then set the FSB to 200 or 266. And if you have any kind of vdroop control make sure to disable it.

Hopefully, this way the multi and voltage will drop automatically to the low VID when idle.

[unclewebb]

rge: You've summed up the stage we're at. Without any further info from Intel, it's about impossible to get any more accurate than what we're at now. Even with more info, there's still variability at both ends of the temperature curve with these sensors so we're still going to be doing some guessing. I think our present guess at the idle delta is pretty close to reality.

[AndrewZorn]

AZ!

Hello. Ninja here.

NO WAY!

[little_scrapper]

I have an obviously stuck Core0 sensor. Yet real temp says its working fine?? What gives?

Core1 reads about 30-31C depending on room temp. Which jives nicely with both core temps BEFORE I moved. But after I moved (company moved us) now my core0 IDLE IS STUCK AT 37c. Load temp is normal and on par, but idle temp is 6-7C higher then core1 and real temp says its responding normaly???



So I have a core0 that wont go below 37C and is rock solid stuck at 37. yet realtemp says its fine

I dont really wnat to RMA it because it is perfectly stable at 4.05GHz @ < 1.28V all day

Last night I cleaned and redid the arctic ceramique paste which made no difference. I also gave it the razor test which gave basically zero light through indicating the IHS was almost perfectly phlat.

[Fungus]

unclewebb, could you maybe add some more font options for the systemtray? a little thicker font like rivatuner has, they are a bit hard to read @ high resolutions due to the thinness.

also, maybe option to have mb cpu ambient temp too?

just some ideas =]