What temp sensor program to trust?

[sebaz]

I have a mild overclock on my i7 3930k, just 4 GHz on Turbo. My motherboard is an Asus P9X79 Pro, so I downloaded the AI Suite. I would think that this being software made by Asus it would show the most accurate sensor readings. But browsing different forums I read that some people say it's unreliable, and it's better to use CoreTemp, Real Temp or one of the other sensor reading programs out there.

What this leaves me with the question, which program do I believe? Of all the people who say AI Suite is unreliable, are they just repeating what they read in forums, or is there an educated reason for saying it's not reliable?

I want to know because I don't care so much about speed as I do about stability and my CPU to last several years. I can't find anywhere what's the real TJMax temp for the 3930k, at least not a reliable source, because on Intel's website it only shows the TJCase as 66.8 and doesn't show TJMax. So depending on what forum you read, some people say it's 91 (which is what CoreTemp and Realtemp show as well) and some people say less than that.

After running Prime 95 for about 25 minutes, AI Suite shows me 58C; Real temp shows me 80, 79, 74, 75, 77 and 80 for each core. So we're talking a 20 degree difference here. Doing more testing with Real Temp and Core Temp both open show that they are about the same, with a 1 degree difference every now and then that is probably more due to update of the readings being off than anything else. AI Suite still lags 20 degrees behind both of them.

So how do I know which program to trust? And how can I find out reliably (not by what people repeat from others) what's the max safe temp for the 3930k that will not shorten its life?

 

[redduc900]

Ai Suite shows the CPU Case temp (read from a thermal diode located within the CPU die between the cores, and what Intel specifies as Tcase), unlike either Core Temp or Real Temp which show the core temp's (read from digital thermal sensors located within each core, and what Intel specifies as Tjunction). Tjunction will always be higher than Tcase, and Tcase always higher than ambient.

 

[NiHaoMike]

First thing I would try is to set sane fan settings. Something like 65C or lower core temp would be good for a 3930k.

The stock fan that is included with a 212 Evo is on the weak side for a 6 core. I suggest a mid range variable speed Delta. 1.5A rated current or so would be good.

 

[sebaz]

Ai Suite shows the CPU Case temp (read from a thermal diode located within the CPU die between the cores, and what Intel specifies as Tcase), unlike either Core Temp or Real Temp which show the core temp's (read from digital thermal sensors located within each core, and what Intel specifies as Tjunction). Tjunction will always be higher than Tcase, and Tcase always higher than ambient.

So if Intel specifies a TJCase of 66.8 for this CPU, and AI Suite is showing me 58C, I should be well within normal limits, right? Meaning, I won't shorten the life of the CPU if AI Suite shows me a temp that stays below 66.8C, correct?

 

[redduc900]

That's correct.

 

[sebaz]

That's correct.

Great, thanks.

 

[unclewebb]

What AI Suite shows is actually some sort of approximation of the TCase temperature. According to Intel, the only way to measure TCase accurately is to cut a groove in to the top of the heat spreader over top of the CPU cores and then you need to solder a thermocouple to the geometric center.

Most end users are not going to hack up their new CPU to properly measure TCase so that spec is completely meaningless. There is no software available that reports the true TCase temperature.

Intel uses core temperature data to control thermal throttling and thermal shutdown. If RealTemp shows in the Thermal Status area that your CPU is OK then that means that it has not reached the maximum safe operating temperature yet and everything is OK.

 

[LoneWolf121188]

Ai Suite shows the CPU Case temp (read from a thermal diode located within the CPU die between the cores, and what Intel specifies as Tcase), unlike either Core Temp or Real Temp which show the core temp's (read from digital thermal sensors located within each core, and what Intel specifies as Tjunction). Tjunction will always be higher than Tcase, and Tcase always higher than ambient.

The problem I've found is that Core Temp, Real Temp, and just about everything else shows a temperature that corresponds a little too closely to CPU loading. Try this: Run Prime, IBT, or whatever else you want, and watch the individual core temps. Then stop the load. INSTANTANEOUSLY (well, within the span of one refresh rate), the temperature drops by 20-30+ degrees! I don't believe that's physically possible. There's no way any cooler can possibly drop the temperature of a descent sized chunk of silicon and aluminum by 20 degrees within a second. This is why I really don't trust any temp program at all.

Just a little math to back me up here: Pure silicon has a specific heat capacity of 0.71 J/(g*K). For ease of calculation, lets say a CPU consists of 20 grams of pure silicon (significantly off, since that's not accounting for the aluminum IHS, which has a way higher specific heat capacity). To cool our 20 gram pure silicon CPU by 25 degrees in 1 second, the cooler would have to have a cooling power of 355 Watts! Even a very powerful peltier can't pull that off!

 

[RGE]

The problem I've found is that Core Temp, Real Temp, and just about everything else shows a temperature that corresponds a little too closely to CPU loading. Try this: Run Prime, IBT, or whatever else you want, and watch the individual core temps. Then stop the load. INSTANTANEOUSLY (well, within the span of one refresh rate), the temperature drops by 20-30+ degrees! I don't believe that's physically possible. There's no way any cooler can possibly drop the temperature of a descent sized chunk of silicon and aluminum by 20 degrees within a second. This is why I really don't trust any temp program at all.

Just a little math to back me up here: Pure silicon has a specific heat capacity of 0.71 J/(g*K). For ease of calculation, lets say a CPU consists of 20 grams of pure silicon (significantly off, since that's not accounting for the aluminum IHS, which has a way higher specific heat capacity). To cool our 20 gram pure silicon CPU by 25 degrees in 1 second, the cooler would have to have a cooling power of 355 Watts! Even a very powerful peltier can't pull that off!

Intels cpus are made of copper laden silicon with a thermal conductance of 180 W/mk per intel. While specific heat will tell you how much watts absorbed for change in temp, thermal conductance is what controls the rate of change. And silicon has a high thermal conductance. If temps did not change instantly from 180W load to 5W idle, I would be suspect, but to each their own.

And Core temp and Realtemp have nothing to do with it. They are simply reading the registers from cpu. So if you dont believe Coretemp and Realtemp you are saying intels digital thermal sensors are incapable of accuracy, in other words intel engineers, and all the engineers at other companies that sell digital thermal sensors capable of 1-5% accuracy (when calibrated for such) are in error.

 

[LoneWolf121188]

Intels cpus are made of copper laden silicon with a thermal conductance of 180 W/mk per intel. While specific heat will tell you how much watts absorbed for change in temp, thermal conductance is what controls the rate of change. And silicon has a high thermal conductance. If temps did not change instantly from 180W load to 5W idle, I would be suspect, but to each their own.

And Core temp and Realtemp have nothing to do with it. They are simply reading the registers from cpu. So if you dont believe Coretemp and Realtemp you are saying intels digital thermal sensors are incapable of accuracy, in other words intel engineers, and all the engineers at other companies that sell digital thermal sensors capable of 1-5% accuracy (when calibrated for such) are in error.

Thermal conductivity tells you the rate of change of energy over a specific distance, hence the meters term in the denominator. In order for that to be a useful figure here, you'd need to know the distance between the thermal probe and the heatsink.

I agree that the thermal conductance is quite high, and I expect that there be little variation in temperature with thickness in a typical CPU. However, if you agree with me that that assumption is true, then the logical conclusion is that all of the heat energy is removed from the whole CPU die in a very short period of time (<1 second, say) when it goes from fully loaded to idle. Hence my specific heat capacity argument: you need to remove ~350 joules from the whole CPU die in that 1 second, hence 350 W. Which seems unreasonable to me.

My conclusion is that the temp monitoring programs are reading the registers wrong. I don't doubt Intel's engineers (I know several personally, and they're all pretty smart folks), I'm doubting a free program I downloaded off the internet.

 

[RGE]

My conclusion is that the temp monitoring programs are reading the registers wrong.

The instant decrease in temps from load to idle has occurred on every past cpu with every temp monitoring program, including when intel supplied its own temp monitoring program on cpus, ie TAT. So again you would be concluding intel's engineers dont know what they are doing, which is nonsense.

Not to mention Intel supplies the information for temp monitoring directly to larger companies like everest, hardware monitor, and those programs read same instant decrease in temp. And Realtemp and Coretemp simply get info directly from intel site, you can go there yourself and read the documentation for MSR and read it yourself with MSR editor.

But either all engineers including intel are wrong and all software including intels in past is wrong or it is just one individual on internet that is wrong...ill let you decide.

 

[LoneWolf121188]

Mind getting me a link to that? I googled around but couldn't find any current info. Again, I'm certainly not saying Intel's engineers are wrong, I'm just trying to figure out how the math works out. Physics is physics.