Hardware Secrets Forums


Go Back   Hardware Secrets Forums > Miscellaneous > Content Comments



Reply
 
Thread Tools Search this Thread Display Modes
Old 12-07-2009, 02:04 AM   #31
Olle P
Senior Member
 
Olle P's Avatar
 
Join Date Oct 2008
Sweden
Posts: 1,736
Olle P is on a distinguished road

Default

Quote:
Originally Posted by Merman View Post
My point is about precision. The developer makes that clear in plain language. There is an error slope ...
I think it's you that confuse precision with accuracy.

If you take a series of repeat measurements at the very same entity, let's say for example a temperature T0, with all surrounding factors and the entity itself identical for all measurements. Then the results will cover some range from Tmin to Tmax, and be close to a bell curve distribution with its peak at the avreage Tavg.
The precision is defined by the difference between Tmin and Tmax while the accuracy is given by the difference between Tavg and T0.

- The error caused by lack of calibration will cause accuracy differences between CPUs.
- The error caused by not knowing what the tables looks like exactly will cause accuracy differences between BIOS/software versions (that use different translation tables).
- The slope error will cause the deviation between measurement average and
actual temperature to increase (in a fairly predictable linear manner) with the distance from Tj.
- None of these factors will increase the spread in readings of a given temperature with the same hard- and software!

Quote:
Originally Posted by Merman View Post
... no one talks about humidity, air pressure and other factors.
I do!

Now going back to the original issue about whether or not a variation in room temperature makes a difference in the heat transfer, as claimed by Halfwaythere, or if it's all proportional to the temperature gradient, as claimed by Raphael:
I suggest that you're both somewhat correct!
- Under the test circumstances one can make the approximation that all transfer of heat is by conductivity from cooling fin to "air" (mostly nitrogen, N2) molecules. That transfer is directly dependant on the initial temperature difference as well as the amount of air molecules.
- The amount of molecules involved is based on the flow rate (caused by the fans) and air density.
- The density is increased when the temperature is decreased, as shown by the general gas law: pV=nRT. R is a defined constant. V is irrelevant and can be seen as constant. p is the air pressure in the room near the coolers air intake, which does vary with the weather, but not much. Remains T, the room temperature, and n, the number of molecules. With the variations in room temperature ranging from 23C down to 14C the difference in density is such that at 14C there are 3.13% more molecules than at 14C, given the same pressure.
- Assuming that cooling performance is directly proportional to the air density it could thus differ by some 1.7% from what it would be with all tests done at 18.5C room temperature, but given that it's only the Intel stock cooler and Cooler Master V10 that's been tested at 14C and thus given a better result than they deserve I don't think it's a big problem.

Cheers
Olle
Olle P is offline   Reply With Quote
new Sponsored Links

Old 12-07-2009, 09:10 AM   #32
Merman
Senior Member
 
Merman's Avatar
 
Join Date May 2008
Posts: 695
Merman is on a distinguished road

Default

Where's your proof Olle??? Intel gives a couple of reasons not to rely on DTS for temperature readings.

Last edited by Merman; 12-07-2009 at 09:14 AM.
Merman is offline   Reply With Quote
Old 12-08-2009, 12:35 AM   #33
Olle P
Senior Member
 
Olle P's Avatar
 
Join Date Oct 2008
Sweden
Posts: 1,736
Olle P is on a distinguished road

Default

Quote:
Originally Posted by Merman View Post
Intel gives a couple of reasons not to rely on DTS for temperature readings.
And I agree!
- You can't use it to read what actual temperature there is!
- You can't use it to compare temperature readings from different CPUs.
- You can't use it to compare temperature readings from different BIOS versions.
- You can't use it to compare temperature readings from different software versions.

My point is that we don't want to do any of that!
- We just want to know the relationship between different temperature readings, using the very same CPU, motherboard, BIOS and software.
- With those limitations you can reliably compare different temperature readings to each other to tell which is higher/lower/equal, even though the given temperature difference between readings isn't necessarily correct in magnitude.

I've attached an image where the blue line represent the true temperature (as taken straight off the X-axis). The magenta and yellow lines represent the corresponding readings from two different CPU/BIOS/software combinations.
No matter what combo is used the relation between two points on the X-axis (true temperature) will be properly reflected on the Y-axis (temperature reading).

Cheers
Olle
Attached Thumbnails
Click image for larger version

Name:	Temps.PNG
Views:	79
Size:	15.9 KB
ID:	194  
Olle P is offline   Reply With Quote
Old 12-08-2009, 02:29 AM   #34
Merman
Senior Member
 
Merman's Avatar
 
Join Date May 2008
Posts: 695
Merman is on a distinguished road

Default

This is all assumed and no one knows the slope error or the individual Tj max, which varies too per CPU. You could probably get more accurate info from a thermometer touching the edge of the CPU.

The only reason software temps seem reasonable are the assumed additions to sensor readings.

Quote:
hendermd: for finding the magic document.

After waiting all these years, I was kind of hoping for more information.

Intel Speak: DTS may ‘bottom out’
Translation: Tough luck buddy, looks like your sensors are sticking.

It's nice of them to publish this but I don't know if we're any further ahead. For my E8400 if TjMax is 100C then the reported temperature based on DTS data would be 47C when the "real" temperature is only 39C. This non linear sensor creates an 8C error here and that error continues to increase as the temperature decreases so at idle I might be out by 15C and my E8400 sensors aren't that bad compared to some I've seen.

If one sensor reads 15C too low and another one on the same processor 15C too high then the data coming from these is pretty much useless for users.

I have no idea where to go from here. It's easy enough to change TjMax but I don't think that brings any of us closer to the actual temperature. Some may get closer but depending on your sensors, you might end up farther away from the real temperature of your processor.

Edit: I've been thinking through this new mess and here's what I've got. Every temperature monitoring program except RealTemp is wrong because it uses a linear model when the data that is coming from these sensors is non-linear. RealTemp is wrong because it is using the wrong TjMax for many processors. Intel's presentation doesn't give us any formulas or anything that anyone can agree on or use to create a formula to convert this DTS data into temperatures.

About all I can do with RealTemp is adjust TjMax and then create the ability to use larger calibration factors to cover up the amount of error that is present in these sensors at idle. Reported temperatures in the temperature range that you operate at may be very close to what RealTemp presently reports if future RealTemp and present RealTemp are both calibrated.

Intel's presentation shows a linear error which I've always assumed but it doesn't clearly show when this error starts. It might start at TjMax.
http://www.xtremesystems.org/FORUMS/...08#post3236208
Unclewebb is the developer of RealTemp. It's a very long thread it documents all the ideas and attemps to use software to monitor (DTS) Intel's CPU sensors. He is very honest and the main point is the sensors are not sensitive enough for tracking temperature. There is too much error and assumptions involved.

The assumptions used by some developers make temperatures and slopes look reasonable.

The new Intel CPUs are suppose to be more sensitive but I haven't followed the disscussion because as long as one doesn't overclock to the point of CPU shutdown the CPU is operating within operating temps whatever that may be. Though as we all know the higher the operating temperature the lower the part's life time.

Quote:
Minimum temperature is useless information on 45nm chips. The sensors are not designed for this and Intel has stated that:

"Any DTS reading below 50°C should be considered to indicate only a temperature below 50°C and not a specific temperature."

At 50C, the amount of slope error means your temps are accurate to plus or minus 10C. Trying to compare that quality of data to other users will get you nowhere fast.

Go read the datasheet for the 45nm Atom processor where Intel tells it like it is concerning these sensors.

5.3 Digital Thermal Sensor
download.intel.com/design/processor/datashts/320032.pdf

For some reason Intel was a little vague at their August IDF presentation and didn't mention specific error numbers for their 45nm Dual and Quad cores processors
http://www.hardforum.com/showthread.php?t=1346091
Another post in that thread:

Quote:
BeastNotro: I can honestly say, I don't know what the real truth is and what's provided in the Intel document is only part of the story. They did not release enough information for any software developer to write an accurate program so we're right back to guessing and making assumptions. It's easy to take a pen and circle some numbers but they didn't test, prove or show anything.

My plan is to start using TjMax=100C for a lot of processors including the E8400 and Q6600 G0.

I looked over some of my old testing and if I now assume TjMax=100C, when my E8400 is showing 47C it is actually at 39C which is an 8C sensor error and when it is showing 29C it is actually only at about 17C which is 12C of sensor error. If I were to use any of the competition's temp software with the correct TjMax, that's how far out they'd be at reporting my E8400.

Both my sensors on this CPU read too high but many 45nm Quads have one sensor that reads too high and one that is too low. Suddenly a combined 20C difference between two sensors only millimeters apart looks about normal. That's a sad situation.

All I can do is follow Intel's guidelines for 45nm and take another shot in the dark at TjMax for the 65nm processors. I will also re-do my calibration formula to continue to give users a way to correct for the significant error that these sensors have. Full disclosure by Intel would have made this easy but once again I'm left guessing. If you've lost faith in RealTemp and this whole mess then I fully understand.

Before I became a programmer I was a user too and I'm not at all happy with this. It's easy for Intel to say that there will be "DTS range and slope improvements" when Core i7 arrives but that doesn't help present enthusiasts one bit. All they're really saying is that the present digital thermal sensors range is very limited and they suffer from significant slope issues where sensor movement is not at all linear with changes in core temperatures.

Quote:
What surprises me is that the extreme quads have a lower TJmax than the non-extremes.
Just one more item that doesn't quite make sense.
http://www.xtremesystems.org/FORUMS/...08#post3236208

Last edited by Cheetos; 12-09-2009 at 09:00 AM.
Merman is offline   Reply With Quote
Old 12-08-2009, 10:50 PM   #35
Olle P
Senior Member
 
Olle P's Avatar
 
Join Date Oct 2008
Sweden
Posts: 1,736
Olle P is on a distinguished road

Default

Merman, out of all your quoted material I've now finally been able to find one tiny bit that contradict my thesis:
Quote:
... the present digital thermal sensors range is very limited and they suffer from significant slope issues where sensor movement is not at all linear with changes in core temperatures.
Everything else has suggested that this is merely a problem when trying to compare readings from different sensors or get the actual temperature.

If the statement in the quotation above is true I agree that it might be less useful to use the integrated thermal sensors when testing CPU coolers.
The magnitude of the problem depends on the magnitude of the errors, especially the non-linear slope.

Cheers
Olle
Olle P is offline   Reply With Quote
Old 12-09-2009, 12:01 AM   #36
Merman
Senior Member
 
Merman's Avatar
 
Join Date May 2008
Posts: 695
Merman is on a distinguished road

Default

To tell the truth I don't know if anyone else is explaining anything about monitoring these sensors.

Curiously AnandTech talked about these problems and then seemed to fall silent after Intel's disclosure (Aug 2008). They also have not had any cooler tests since before the 2008 disclosure. Jan 2008 was there last cooler tests. I am not trying to start any conspiracy theory but it seems to me the subject was dropped by developers and Intel pulled its report from the net.

http://www.anandtech.com/weblog/defa...x?bcategory=12

Quote:
We compare cooling results with a representative sample of air and water-cooling results measured with CoreTemp. TAT provides a similar core measurement, but test results with CoreTemp are more consistent over a wide range of test conditions than the results reported by TAT. We retested previously reviewed coolers with CoreTemp under idle and load conditions. In benchmarks where the new test bed makes no apparent difference, like maximum overclock, we include results for all coolers tested since beginning cooling reviews in early 2007.
http://www.anandtech.com/casecoolingpsus/showdoc.aspx?i=3210&p=5
Here is a link to the Intel document in HTML format.

View as HTML


[PDF]
Verdana Bold 30

PDF may show the graphs better as that is the document's original format.

Last edited by Cheetos; 12-09-2009 at 08:57 AM.
Merman is offline   Reply With Quote
new Sponsored Links

Reply

Share This Thread

Bookmarks

Thread Tools Search this Thread
Search this Thread:

Advanced Search
Display Modes

Forum Jump


All times are GMT -8. The time now is 05:13 PM.


vBulletin Security provided by vBSecurity (Lite) - vBulletin Mods & Addons Copyright © 2014 DragonByte Technologies Ltd.
vBulletin Optimisation provided by vB Optimise (Lite) - vBulletin Mods & Addons Copyright © 2014 DragonByte Technologies Ltd. ()
© 2004-12, Hardware Secrets, LLC. All rights reserved.