i7 950 MAX Overclock on air

[bmg]

For the record, LOAD LINE CALIBRATION WAS VERY IMPORTANT! Night and Day difference of stability.

From your earlier posts it sounded like you had it disabled. You're getting more stability with it enabled or disabled?

[Nanometer]

It seems like it. My only change has been enable load line calibration, which seems to raise vcore slightly underload. No complaints here. I'd rather vcore increase under load then when it's at idle. 1.38V is about what I need for 4.4GHz. 4.6GHz is only going to happen on a cold night, at least stable wise. I'm running high volts now, but I'm still hitting 82C under load with my case closed.

HT is disabled since it adds no performance to single threaded apps. under 70c was the load, with the case open with several fans blowing cold air, about 62F air at the case. No promises of proof. But I do plan to have more fun later and I will take screen shots. The pictures haven't changed much, just that the window is open and theres several fans pulling cold air to the open case.

Also one of the big changes I made a few days ago was adding the Xigmax or whatever the name is chipset cooler to the northbridge set. dropped NB temp from 75 to 55 under load. Though I still had stability after that installation, I feel better with the chipset at a lower temp.

Once again loadline calibration seemed to be a big issue for stress testing. Enable it, all it seems to do is add vcore underload, or at least stable it out slightly. loadline calibration is a dream when it comes to vdroop. no more ic7maxIII or p4c800 vdroop mods here

Also for the record. PCI-e frequency above 110 seemed to give terrible errors and really screwed up the system. I'm pretty sure it lead to the numerous SSD errors I had. So I had to reformat. Maximum PCI-e I would use would be 110 anything higher with this board and you are asking for trouble.

Another thing. I've been having cold boot problems as well. Sometimes it takes a few attempts to boot up, though at the moment my current setup is at least 30 minutes prime stable.

[TheCarLessDriven]

Why do you need to type in bold?

[Leeghoofd]

What was your LLC setting before ? AUTO or Disabled ? as AUTO is the same as enabled on my board (0809 Bios) , it adds a bit more Vcore when under load (compared to the bios setting)

[Gaul]

oh man, daily 4.4 gh thats very cool on air

[Nanometer]

I believe LLC was set to disabled, hence about .05 drop in voltage.. with it enabled it increases to about .02 i havent really payed attention to auto.

yea, 4.4 is very nice

ram wont run 6-6-6-x best i can do is 7-7-7-18 at 1600mhz. So i bumped it up to 9-9-9-24 2000mhz for now, i might be able to tightening some timings. Only difference I have seen going to 2000mhz is increase in memory bandwidth all over the board, no surprises. I'm still looking for minimum 4.4 voltage.. so it might be a while before i get final results. With that said, I wont be going for 4.6, not even for suicide shots. If I had water, I might, but at 4.4ghz and a fully balanced system, i am guessing i will be loading at 80C on the hottest core which is where i want to be. For some odd reason core 3 is 9 degrees cooler then core 0. about 71c.

I also wish this boardgave the option for independent multipliers so I could set one to 4.6ghz for single threaded apps... oh well

max 247 settings i will do will be 4.2 because of temperature under prime. 4.6 just isnt happening 247 on air. 4.4 could be done, but then we are talking 90c+ load temps.. which im not gonna do. id rather see 82c and less.

[Nanometer]

So I got fed up and put my watercooling back in. Pictures speak for themselves, but keep in mind that actual vcore is 1.21v at the meter



thanks about the OC but big thanks to this pretty nice chip. im taking her up now with HT enabled. id be thrilled with 4ghz but i dont think ill get it, at least not with 240 radiator.. it gets hot

[espionage]

whoa, that is one hell of an oc. I must get one of the D0 now for my new i7 setup.

[nyeah]

sorry,but could you post a bigger SS ?
i couldn't see it on my monitor
nice result btw

[Nanometer]

It's 4.06GHz now at 1.19volts HT disabled. I'm now going up with HT on.it's difficult to post pictures since the forum wont let us post pictures using this site as an archive.

[CompuTronix]

Guys,

I realize this thread is over 4 months old, however, a forum member at Tom's brought this to my attention, so I'd like to clear up a few misconceptions.

Just make sure none of your cores for your 950 pass 85C and you're good.

Intel has no such 85c thermal specification for any Core i7 variant. Please refer to Intel's Processor Spec Finder - http://processorfinder.intel.com/Lis...052&SearchKey= - and Intel's Core i7 9xx Datasheet - http://download.intel.com/design/pro...hts/320834.pdf

And you mean core temp not socket temp (CPU) right?

The erra of the old thermocouple-in-the-socket-touching-the-underside-of-the-CPU became history before single cores became dinosaurs. The Analog Thermal Diode (CPU temperature sensor) is integrated within the processor package substrate, dead-center in the lower layers. Please refer to the following Intel document: http://arxiv.org/ftp/arxiv/papers/0709/0709.1861.pdf

Guys,

For everyone's benefit, here's what you need to know:

Temperatures and overclocking are all about specifications, so it's very important to be specific. If we're not, then the topic makes about as much sense as comparing apples-to-oranges thermal fruit salad in a blender! My objective is to assure that enthusiasts understand Intel's specifications, standards and test methods, so they can better decide how to apply and manage their overclocking options.

From Intel's Processor Spec Finder - http://processorfinder.intel.com/Lis...052&SearchKey=

All Core i7 9xx variants:

Vcore Max 1.375v
Tcase Max (CPU temperature) 68c
Tjunction (Core temperature) 73c

From the Core i7 and Core 2 Temperature Guide - http://www.tomshardware.com/forum/22...perature-guide


"Section 1: Introduction

Core i and Core 2 processors have 2 different types of temperature sensors; a CPU case (not computer case) Thermal Diode centered under the Cores, and Digital Thermal Sensors located on each Core. The case Thermal Diode measures Tcase (Temperature case), which is CPU temperature, and the Digital Thermal Sensors measure Tjunction (Temperature junction), which is Core temperature. Since these sensors measure 2 distinct thermal levels, there is a 5c temperature difference between them, which is Tcase to Tjunction Gradient. Core i7’s / i5’s and Core 2 Quad’s have 1 Tcase and 4 Tjunction sensors, while Core 2 Duo's have 1 Tcase and 2 Tjunction sensors ...

... The monitoring utilities provided by motherboard manufacturers monitor CPU temperature, while some popular freeware utilities monitor Core temperatures ... Real Temp ... is recommended for users interested in monitoring Core temperatures only ... SpeedFan monitors Tcase (CPU temperature) and Tjunction (Core temperature) ... "


The Thermal Specification shown in Intel's Processor Spec Finder is Tcase Max (CPU) not Tjunction (Core), which is a very common misconception among most enthusiasts. Since there's a 5c gradient between the CPU sensor and the Core sensors, (shown in the following Intel document) - http://arxiv.org/ftp/arxiv/papers/0709/0709.1861.pdf - just add 5c to the value shown in the Spec Finder to determine the corresponding Core temperature, which is 73c for all Core i7 9xx variants.

Intel's second and frequently misunderstood Thermal Specification, Tjunction Max, (100c for all Core i7 9xx variants) applies to overtemp protection such as Throttle and Shutdown, so you don't toast your transistors. As such, any i7 9xx Core temperatures which exceed 73c should be considered "overtemp". Further, when specifications are exceeded, then processor degradation becomes a concern, which is explained in the following AnandTech article - http://anandtech.com/cpuchipsets/int...spx?i=3251&p=6

Prime95 Small FFT's is the Standard for processor thermal testing, because it's a steady-state 100% workload which yields steady-state temperatures, whereas Blend is a memory cyclic workload which yields fluctuating processor temperatures. Small FFT's will reach 97% thermal saturation within 7 to 8 minutes, so a 10 minute test is adequate. Thermal testing should be conducted as close as possible to 22c (72f) Standard ambient, with case covers removed, the computer clear of any desk enclosures, and all fans at 100% RPM to eliminate cooling variables, and to produce consistent and repeatable results for comparisons. If the Gradient between CPU temperature and "mean" (average) Core temperature is not ~ 5c, then BIOS is incorrectly coded. CPU temperature and Core temperatures can be individually calibrated in SpeedFan by following the Calibrations Section in the Temperature Guide.

OCCT and Burn Test (reminiscent of TAT) use LinPack, which shows thermal signatures that resemble the ups and downs of a bad day on the stock market, and cycle between light workloads, through test segments which spray all processor registers with all one's, (100% thermal load = 115% workload), and can push an overclocked i7 9xx with HT enabled at Vcore Max 1.375, right on past Tcase Max to ring the Tjunction Max bell like a fire alarm!

Since there are very few applications or games that will spike, let alone sustain processor workloads beyond 70% to 85%, utilities which load all registers with all one's are not representative of real-world computing. While these utilities are certainly very useful for stability testing, they are inappropriate for thermal testing. The 3DMark benches are excellent for stability testing, as are applications for ripping and encoding.

To make sense of CPU temperature and Core temperature, compare them to a 4 cylinder car with 5 temperature guages; 4 of the guages are cyclinder head temperatures (closest to the heat source), and the 5th guage is the overall engine temperature, which is 5c lower than the other guages, and is the temperature guage with which we're all familiar. We know the red zone (hot) for the i7 9xx starts at 68c (Tcase Max) on the engine temp guage and 73c (Tjunction) on the cylinder head temp guages, but if we push the engine too hard and peg all the guages, (95c Tcase overtemp / 100c Tjunction Max) then the engine will shut down.

If you'd like to learn more about processor temperatures, then just click on the link in my signature.

Hope this helps,

Comp

[unclewebb]

Hi Comp,

I had a look at the Intel documentation you posted and I saw the chart that shows the difference between the peak TCase diode temperature vs the peak core temperature. There is no temperature scale on the left hand side of that chart. I don't agree with your assumption that the scale is 1C per division. I think the scale is more likely 5C per division. If that is so then maybe the graph shows that there can be as much as a 26C gradient between TCase and core temperature depending on what program you're running.



That would make a lot more sense and would agree with the testing that rge did when he drilled a hole in the middle of his Core 2 CPU and mounted a calibrated thermocouple in the geometric center like Intel recommends for monitoring TCase temperatures. I think when he was running Prime95 he was seeing gradients between these two temperatures in the 25C range.

I'll send him a PM and a link to this thread so hopefully he can share some of his testing. He might still have a link to the YouTube video he did when testing.

Here's a quote from that Intel paper:

"It can be seen that large temperature gradients exist on the die. It also can be noted that some workloads display high temperature gradients while other have no offset. Thermal control algorithms need to prevent the hot spot from exceeding the max temperature specification. It is possible to mitigate the temperature difference by applying a fixed offset to the diode reading. This obviously is a non optimal solution as the workloads with low offset will be penalized by the unnecessary temperature offset. The use of digital thermometer provides improved temperature reading, enables higher CPU performance within thermal limitations and improves reliability."

My interpretation of that paragraph is this. Intel introduced core temperature sensors so they could better control thermal throttling and thermal shutdown without having to penalize users with early thermal throttling based on the inaccurate TCase diode temperature. Thermal control in modern Intel CPUs is controlled by the core temperature sensors to provide users with maximum performance without having to worry about thermal throttling until absolutely necessary based on the hottest spot on the core.

[rge]

There are three temps:
1) Core temp in die via DTS (i7 tjmax 100C)>>>gradient over small distance of die substrate to the>>2) cpu diode (still in die substrate)>>> 3) then gradient all way through the rest of die substrate to top of die, across solder tim, across IHS to ACTUAL CPU/IHS temp (68C Tcase max i7) which is published intel specs.

The gradient between Core temp and actual IHS temp on core 2 duo or pentium M at ~65W TDP varies from 8-10C on load with minimal loading programs, to 15+C with prime to 22+C with Linx, and that is at stock. AT overclock settings the gradients will be higher (higher TDP). That I have videos of with drilled holes and thermocouples, and that was also tested by stanford university with 22 different loading programs, and intel has published formulas stating such. And it is this gradient up to 32C in I7 with stock settings/cooler that is difference between core temps (tjmax 100C) and cpu temps (68C). However intel basically states you can exceed core temp specs without exceeding TCASE specs, since gradient varies, though it is difficult to do with adequate cooling. For example a 20C gradient loading program at 92C core would be 72C Case, but difficult to accomplish with good cooling.

Then comes the mess. CPU diode is located in die substrate but calibrated by bios per intel specs to read much lower than its actual location, so it approximates IHS temp. For example, when my core temp reads 85C, cpu diode temp calibrated by bios reads about ~60C. And in fact if measured IHS with thermocouple, temp would be somewhere around 60C if calibrated properly. However it is correct in stating actual temp in die substrate at actual diode location will not be 60C, but will be higher. But again, intel is not using this cpu diode to measure that location, since that temp at that location is not relevant. The cpu diode is purposely set low to approximate IHS temp. And if that cpu diode is then calibrated to read its true location, it will then read much higher than IHS temp, will no longer represent IHS temp, and hence becomes an irrelevant temp in die substrate, which is too low for hot spot temp and too high for cpu/IHS specs temp.

As for the gradient between core temp and cpu diode, I am not 100% sure in 65nm scale (not relevant on i7s anyway with 2xTDP and different die size/power density), and dont think it matters, as the gradient that matters traverses from die diode to cpu diode then continues all way through die substrate to top die, across tim1 and IHS. I am 100% sure max avg spec at stock gradient from tjmax (core) to Tcase max (IHS) is 32C on I7, ie tjmax-tcasemax.

But if your core i7 is at 85C, with adequate cooling, IHS temp is likely to be lower than 60C, so not exceeding either tjmax or tcase max. As to the temp in die substate, whether it is 5C or 10C or whatever, I dont know. I do know the gradient across the entire die substrate can be 20+C, but dont know the proportional distance from die diode to cpu diode versus die diode to top of die.

[CompuTronix]

unclewebb and rge,

Always a pleasure!

Although I seldom offer comments on the Real Temp thread, I have followed every page, and continue to do so. As such, I am also aware of your excellent work, rge, and I agree with your findings completely.

While I see your point concerning how the paragraph is worded, I believe that Figure 5 "Diode to DTS Temp. difference" refers to the Analog Sensor shown in Figure 1, rather than the Thermocouple / IHS "Tcase" measurements detailed in Intel's Thermal Metrology, which is not discussed in this document. However, considering the lack of a defined scale, it does leave things open to interpretation as to exactly what a "large temperature gradient" means in relative terms.

I realize and appreciate that the scope of your efforts are focused primarily on Core temperatures, however, many enthusiast are continually confused by CPU temperature, and it's significance relatative to Core temperatures, especially when they encouter mis-programmed BIOS, CPU temperatures below ambient or above Core temperature, or a "stuck" DTS. Nonetheless, CPU temperature, when properly calibrated, provides an additional thermal reference which can be quite useful under a number of scenarios.

The following is from post #3531 of 6/16/09 on page 142 of the Real Temp thread:

Hi unclewebb and rge,

Sorry I haven't posted or PM'd for quite some time, however, I've continued to follow your excellent work, as well as every post on every page, every day.

Your points are well taken concerning Intel's deviations in Tjunction Max values, however, from a different perspective, I'd like to share with you my research and empirical data, which I've continued to acquire for my Core i7 and Core 2 Temperature Guide over at Tom's.

During the past 2+ years, I have extensively tested, and recently re-tested and analyzed processor temperatures using three data points which consist of idle, 50% load and 100% load at stock settings with Prime95 Small FFT's. I then repeated each test at overclocked settings with higher Vcore and temperatures. The objective was to observe and carefully document the thermal relationships between CPU temperature and Core temperature, sensor linearity characteristics, slope error behaviors, and power consumption.

Calibrations were based upon the 5c thermal Gradient between the Analog Thermal Diode and the DTS value, of which we're both familiar from certain Intel engineering documents. All testing was conducted using a standardized test setup under controlled conditions at 22c ambient on a variety of motherboards, chipsets and CPU cooler combinations. Real Temp was used to cross-reference Core temperatures in SpeedFan.

Although the CPU temperature offset calibration value may shift on certain motherboard when BIOS test settings are changed, CPU temperature was verified or re-calibrated in SpeedFan prior to acquiring each measurement, so that information from the Analog Thermal Diode would provide a valid and accurate point of reference. This assured that the results were both consistent and repeatable.

The following assortment of 65nm an 45nm processor variants were tested:

(1) E2160 L2
(1) E4500 M0
(1) E5200 R0
(1) E6400 B2
(4) E6600 B2
(1) E6850 G0
(6) Q6600 G0
(1) Q6700 G0
(1) E7200 M0
(1) E8500 E0
(3) Q9650 E0
(6) i7 920 C0
(1) i7 940 C0

Of these 28 individual processors, 17 of which are quads, my findings show that the Analog Thermal Diode is typically linear from low idle temperatures thru very high load temperatures. Testing also revealed that the 5c Gradient between CPU temperature and "Mean" Core temperature is relatively constant from 4c to 6c, which conforms with the detailed thermocouple testing performed by rge.

The most interesting point is that when you originally conducted your IR testing, except for a few low order variables concealed from your Fluke IR gun on the IHS, apparently you were closer than you may have since thought. For example, the Core 2 processors which I've tested that are suposedly Tjunction Max 100c have shown values that average from 97c to 98c, with the exception of the i7's, which are nearly spot-on between 99c and 100c.

As such, it's likely that the differences often seen in the Core 2 quads between each pair of dual cores is 100c and 95c, which yields a Mean Tjunction Max value of 97.5c. This obviously coincides with my findings, and offers further insights toward Intel's new term "Tj Target".

Comp

After nearly three years of work on this topic, it's difficult to consider that I've somehow misinterpreted the fundamental principals. Come to think of it rge, you and I had this CPU temp versus Tcase discussion at Tom's probably a year and a half ago. Regardless, whether the the CPU to Core gradient is actually between 3 to 7c among the processors we've collectively sampled, we're down to splitting fine hairs on this, because we're all within a few degrees.

Comp

[unclewebb]

I apologize Comp, I was misinterpreting the analog diode temperature with the IHS geometric center temperature.

According to Intel, the 67.9°C Thermal Specification for a Core i7-920 refers to the maximum Tcase temperature.

Thermal Specification: The thermal specification shown is the maximum case temperature at the maximum Thermal Design Power (TDP) value for that processor. It is measured at the geometric center on the topside of the processor integrated heat spreader.

In my opinion, this specification is directed at system builders that have the necessary resources to cut a groove in their IHS and correctly mount a calibrated thermocouple. That specification is not directed at the average end user for that reason. If an end user can't accurately measure Tcase then there's no easy way to determine if they are within this specification or not.

All Core i7 9xx variants:

Vcore Max 1.375v
Tcase Max (CPU temperature) 68c
Tjunction (Core temperature) 73c

Here's where I disagree. The Tcase specification temperature is not equivalent to the CPU temperature as reported from the Analog Diode. Your extensive testing has shown a 5C difference between the Analog Diode and the Tjunction based core temperature. In your example, you are now applying this 5C difference to the Tcase specification but your testing was based on measurements from the Analog diode. Tcase and Analog diode are not the same so this comparison or your recommended maximum Tjunction temperature is not valid.

With the introduction of core temperature digital thermal sensors, I don't think the data from the Analog Diode (what SpeedFan reports as CPU) has any relevance anymore. That reported temperature is not equivalent to a correctly measured Tcase temperature so you can't use this data to see if you are within the Thermal Specification. The CPU itself also does not use information from this sensor to determine thermal throttling or thermal shutdown. Both of those are determined by the core temperature.

Intel says:

Regardless of the configuration selected, PROCHOT# will always indicate the thermal status of the processor.

When the PROCHOT# signal is activated, the CPU will start to thermal throttle. After reading through a mountain of Intel documentation, that seems to be the only important thing to monitor or be concerned about.

If PROCHOT# has not been activated then your CPU is running at full speed and within the Intel design specification. That's why I included the reporting of that bit separately in the Thermal Status area of RealTemp. If this bit has not been set since power on, RealTemp will report OK. If your CPU gets too hot and the PROCHOT# signal is activated for even a millisecond, RealTemp reads this bit and will report it as LOG. That shows that at least one thermal throttling episode has been logged. If throttling the CPU by lowering the multiplier and lowering the core voltage isn't able to control the temperature of the processor then thermal throttling will be happening full time and RealTemp will report this as HOT in the Thermal Status area.

I think this quote really sums things up.

Thermal and Mechanical Design Guidelines
August 2008

4.2.6 System Considerations

A system designed to meet the thermal profile specification published in the processor datasheet greatly reduces the probability of real applications causing the thermal control circuit to activate under normal operating conditions. Systems that do not meet these specifications could be subject to more frequent activation of the thermal control circuit depending upon ambient air temperature and application power profile.

Moreover, if a system is significantly under designed, there is a risk that the Thermal Monitor feature will not be capable of reducing the processor power and temperature and the processor could shutdown and signal THERMTRIP#.

Intel wants computers designed so they are not tripping PROCHOT# or activating the thermal control circuit during normal usage so a user will be able to get the full rated speed out of their CPU.

The Tcase specification is designed to guide system builders. If they can design a computer and keep the measured IHS temperature below the Tcase specification then it is very unlikely during normal usage or even when running the most demanding applications like Prime95 or Linpack that the core temperature will ever get hot enough to trigger PROCHOT# and thermal throttling. If your measured Tcase temperature is 68C then even if a user is running Linpack and his peak core temperature is 30C beyond his Tcase temperature, the CPU will continue to operate at full speed and within specification because PROCHOT# will not have triggered.

(95c Tcase overtemp / 100c Tjunction Max) then the engine will shut down.

Just to be clear, a core temperature of 100C is not the thermal shutdown temperature (THERMTRIP#) of a Core CPU. The thermal shutdown temperature is approximately 25C beyond the typical 100C maximum junction temperature so thermal shutdown usually doesn't occur until a core temperature of 125C or even a couple of degrees higher during testing.

Thanks for explaining the Analog Diode / CPU temperature. I've never found this sensor very useful on my motherboard because of its calibration but now that I understand it better, maybe it will make some more sense.

[CompuTronix]

I should've said "the engine will throttle down".

I thoroughly understand the distinction between Tcase and CPU temp, as well as all you've mentioned, since we've probably read most of the same documents. While I agree with you on several points, let's move directly to practical applications to work this out. This discussion might be better served by example, so perhaps if I show the calibrations on my personal rig performed according to my Guide, then you'll see how the 5c gradient makes sense.

Note 1: Core 0 tests as the most linear Core from idle thru load at any BIOS settings.
Note 2: All testing was performed at 22c Standard Ambient.
Note 3: All Vcore measurements are from CPU-Z.

SpeedFan
CPU Offset: -6
Core Offsets: -3,0,1,3 (Mean 100.25)

Real Temp
TJmax: 97,99,101,102 (Mean 99.75)
Idle Calibration: 0.0,0.8,-3.2,-0.8

Test Setup
Case Covers = Removed (Antec 900 - immaculate cable management)
Case Fans = Manual 100% RPM
CPU Fans = Manual 100% RPM (Xiggy 1283 mod'd for push-pull 74CFM each)

BIOS = Stock (Full Auto)
Idle = 10 minutes < 1% CPU Usage
Vcore = 1.056
CPU 24 (<-- RGE's findings at low power with high-end push-pull air)
Cores 28

BIOS = Stock (Full Auto)
Load = 10 minutes Small FFT's
Vcore = 1.208
CPU 46
Cores 51

BIOS = OC'd (Full Manual @ 4.1Ghz HT Off)
Idle = 10 minutes < 1% CPU Usage
Vcore = 1.360
CPU 31
Cores 36

BIOS = OC'd (Full Manual @ 4.1Ghz HT Off)
Load = 10 minutes Small FFT's
Vcore = 1.360
CPU 61
Cores 66

Additional overclock settings and temperatures:

3.0 HT Off - Vcore 1.000
Idle: CPU 24, Cores 28
Load: CPU 35, Cores 40

3.5 HT On - Vcore 1.160
Idle: CPU 26, Cores 31
Load: CPU 49, Cores 54

3.8 HT On - Vcore 1.304
Idle: CPU 29, Cores 34
Load: CPU 64, Cores 69

4.0 HT Off - Vcore 1.304
Idle: CPU 30, Cores 35
Load: CPU 55, Cores 60

4.0 HT On - Vcore 1.432
Idle: CPU 32, Cores 37
Load: CPU 84, Cores 89

4.2 HT Off - Vcore 1.424
Idle: CPU 32, Cores 37
Load: CPU 69, Cores 74

This is empirical data. The temperatures meet all the requirements. These results are consistent, repeatable, typical and plausible. Since the Cores are calibrated to preserve mean TJ Max 100c, these results should be valid. Would you calibrate the Cores, or CPU, or both, differently to find the "Real Temp"?

We've all invested hundreds of hours reading, testing and obsessing temperatures, and have accumulated considerable knowledge, while our buddies were fishing. I've focused on analyzing the interactions between CPU and Core temperatures. If there's a fly in the ointment with using a 5c gradient, then please point it out.

[rge]

What Unclewebb is saying, is the same discussion we did all have before. Intel specs of 68C Tcase refers to IHS temp, not the cpu diode temp. If cpu diode temp is calibrated at 130W TDP using stock, intel listed variables to read 68C when core temp reaches 100C, then it is calibrated for that one load and TDP and cooling to be cpu specs. But if calibrate it to read 5C away from core temp, it becomes meaningless for intel thermal specs at Tcase. And since CPU temp is rarely going to be accurate, except for one particular TDP, cooling, mobo, etc, it is always going to be significantly less accurate than core temps, calibrated or not.

Most engineering articles will state IC chips should not be run for extended times above 100C, and never above 120C, so not hard to understand intel wanting to measure hot spots and correspondingly set throttling and shutoff. Prior to core sensors, the only way to know if you were running at 100C was assuming gradient through chips were all similar and making guestimates with either IHS temp or with cpu diodes, neither were as accurate as core temps now, however.

So to me, with intel new i7's processors that have fairly accurate core temps even in idle ranges, cpu temp becomes irrelevant. If you look at notebook avg temps, many people, including couple notebooks at work, have core temps in high 80s to 90s constantly for years without issue, some notebooks simply run very hot, never cleaned, etc. While 90's is not my cup of tea, no one has degradation curves to really no what core temp is cut off, other than intel saying cpus are not meant to run near throttling for extended times. It is up to user to guess how far away from throttling.

The only way cpu temps would represent IHS specs, would be to measure IHS and calibrate at one load, or to guestimate at set distance from core temps ie like 30C using stock cooler at max load. But soon as OCing, TDP changes and gradient changes.

Bottom line, intel has two specs. Tjmax of 100C and IHS max of 68C on i7. And these are same specs, ie tjmax of 100C is when IHS = 68C, at set TDP, stock, etc. So 2 reasons to disregard IHS temp, one you cant measure it, 2 it is less relevant than actually using the core temp, now that intel has found way to accurately measure hot spots. And cpu diode temp does not accurately represent either, no matter how it is calibrated. All you can really say is keep tjmax from 100C and that is same thing as saying keep IHS temp from 68C at given TDP,etc

Edit: what is really needed is mttf curves at different volts and temps if one wants to guess at appropriate temps at Ocing volts. Intel had published in past at a now dead link (http://www.intel.com/design/PACKTECH/letter.htm) that the internal goal was less than 1% failure rate of cpus at 7 years, and less than 3% failure at 10 years, and this is for stock settings with stock intel cooler. That included random failures from latent defects and early end of life failure. Which explains why many laptops, despite running near throttling temps, have no issues lasting many years. Even if you start with normal loading temp of 70's and double the mean time to failure for every 15C up to 100C, still less than 4 out of 100 cpus die in 7 years, of course that does not take into account the bigger killer, insane volts. I am primarily interested in lowering my temps for stability. At 4.4 ghz, temps dont mean much, I am stable at 90C or 68C loading temps. But on fringe at 4.6 or 4.7 benching, every 10C dramatically helps stability. As for whether temps of 70C loading vs 95C loading helps slow degradation on one particular highly overclocked cpu (temps to me are irrelevant at stock volts since I wont have it for 10 years), again would love to see mttf/degradation curves at different temps and volts to calculate probability curves...but intel wont be sharing that info anytime soon.

[unclewebb]

Since the Cores are calibrated to preserve mean TJ Max 100c, these results should be valid.

Intel does not adequately document TJMax, even in their new Core i7/i5 processors where this value is written into each core. Sure, most Core i7-900 series processors have 100C written into each core but no one really knows what that means. There's no documentation to show how much error is in this value and no documentation to say whether TJMax = 100C represents a minimum, maximum or average value. Your guess is no better or worse than my best guess because there isn't any publicly available documentation to back our opinions up.

I totally agree with your 5C findings that what SpeedFan reports as the CPU temperature as read from the thermal diode should be 5C lower than the core temperature. The problem is that whatever this thermal diode temperature reads, no matter how accurately you are able to calibrate it, that is still not a measurement of the Tcase temperature. You can't compare a thermal diode reading to the Intel Tcase specification because these are based on two different measurements at two different spots on the CPU.

On my motherboard, I can't calibrate this sensor using SpeedFan because as the temperature increases, the amount of error in this sensor also seems to increase. At TJMax when the core temperature is being reported at 98C, the CPU sensor is reporting 107C. Based on your 5C results, it should be reporting about 93C which is a 14 degree error. At a lower temperature where the slope error of the core sensors is still minimal, it reports 72C when the core temperature reports 67C. Based on your 5C findings, the CPU sensor should be reporting 62C so now the error has decreased to 10 degrees. At idle, the amount of error in this sensor has decreased again and now the error is somewhere around 6C.

I don't know if this shows a problem with this sensor or more likely just the calibration correction factor that Asus decided to apply to it. Using a single point offset like SpeedFan allows you to do can make this reading more accurate but no matter where you choose to calibrate it, the readings at other points along the temperature curve will either be too high or too low and not 100% accurate.

That leaves me with a sensor that can't be properly calibrated that displays a number which Intel has published no specification to show whether that number is good or bad. I can now see that if your core sensors were damaged or stuck then you might be able to use data from this sensor to approximate your core temperature as long as your motherboard has applied an appropriate calibration correction to it. In reality though, almost all core sensors work perfectly fine in the 70C to 125C range where they were designed and calibrated to be reasonably accurate. The CPU diode temperature might be useful for users with stuck sensors but sticking sensors at normal temperatures isn't an issue anymore for the Core i7/i5 CPUs.

Intel saw the limitations of using the analog diode sensor which was their motivation to introduce core temperature sensors. It makes a lot more sense to thermal manage a CPU based on information coming from well calibrated sensors mounted on the hottest spots on the core. Even 45nm Core 2 sensors that have a list of issues as long as your arm still work excellent in this upper range. TJMax is not properly documented so using them to report an accurate absolute temperature is a crap shoot at best. They are perfectly fine though to compare your core temperatures from one day to the next and to determine if changing your heat sink or fan or thermal paste improved your full load temperatures or not as long as they are not sticking.

My findings are the same as what rge has found. Core temperature isn't that important until you are overclocking your CPU to the very limit of what it is capable of. At that point, for stability, lowering your core temperature as much as possible becomes very important.

I've spent a pile of time too investigating CPU temperatures and my conclusion is simple. Run your Intel CPU as cool as possible and you will increase your stability while maximizing the amount you can overclock it. As long as your computer is stable and not setting off the PROCHOT# bit and thermal throttling, then you can pretty much ignore your CPU or core temperature. I hate to say it but CPU temperatures just aren't that important anymore. Intel makes some great CPUs that do an excellent job of looking after themselves, even when pushed to the outer limits so there's really no point in users splitting hairs trying to come up with 100% accurate idle temperatures. It's been a nice hobby and I've learned a lot but it's time to move on.

[CompuTronix]

The experiences with your P5B are the exception rather than the norm. I've found that CPU:Core temperatures typically scale 1:1, and the analog thermal diode is as linear as the best core. CPU temperature does only what it was designed to do; provide a measurement of the overall temperature of the entire processor package.

We know that CPU temp must be higher than ambient but lower than Core temp. At stock BIOS settings, processor idle power has known values, just as coolers have known dissipation values, so we can calculate what the CPU temp should be. Give or take a few variables, this works well enough to get us certainly within a degree or 2, so in my opinion, CPU temp is a useful measurement.

Regardless, it's very clear that I'm spinning my wheels here, and that I've wasted a tremendous amount of time over the past three years persuing practical solutions. Henceforth, my recommendations to all will be to ignore CPU temp, and just use Real Temp. My only regret is that utimately, I couldn't contribute anything more viable than a seemingly misguided Guide.

And you're right ... it's time for me to move on. Thanks for the wakeup call.