Delidded my i7-3770K, loaded temperatures drop by 20°C at 4.7GHz

[HondaCop]

An 8C improvement is nothing to sneeze at. So you wonder what the voltage requirement might be for 4.7 Ghz.

Bonzai, I'm running LinX again while I type this...

4.8GHz @ 1.31v

Right now it's done 11 of 20 runs and the temps are:

Core 1 - 57.4C
Core 2 - 58.3C
Core 3 - 64.0C
Core 4 - 57.8C

The highest temp reached so far, was Core 3 and it's maxed at 70C! It's been low to mid 60 through out the test so far. I'm am extremely pleased!

This is insane, guys. Earlier this morning, I was trying 4.8Ghz and at 1.33v, I was averaging mid to upper 90s with MX-4! Unreal!

 

[moonbogg]

As for the shim, I already ordered a roll of 0.203mm metal shim from the supplier linked in that earlier post by yottabit. Will probably be a week before I get it though

That is hardened steel. Its your choice of course, but copper is much more thermally conductive, much softer and less likely to increase risk of cracking the die. Could you cancel the order and go with copper instead? Just a suggestion. You never see steel being used for the base of heat sinks for the poor thermal conductivity. Steel really holds onto the heat. It might act as an insulator.

 

[HondaCop]

LinX test (21000 | 20 runs) with a voltage of 1.40v @ 4.9GHz:

Core 1 - 60.6C
Core 2 - 63.4C
Core 3 - 68.2C
Core 4 - 61.3C
CPU package - 67.8C

I have no doubt in my mind that the thermal issues with Ivy Bridge is due to the TIM route Intel decided to take with these CPUs, instead of soldering them like they used to do. If you really want to get the most of these processors and overclock them without burning the house down, you definitely need to look into de-lidding them and replacing the TIM with a high quality one. Obviously, I will recommend Coollaboratory's Liquid Metal Ultra based on MY findings. The past few weeks I have been trying like a madman, to get a stable overclock of 4.9GHz, but the temps were through the roof in the upper 90s on full load with the same 1.40 voltage.

 

[moonbogg]

LinX test (21000 | 20 runs) with a voltage of 1.40v @ 4.9GHz:

Core 1 - 60.6C
Core 2 - 63.4C
Core 3 - 68.2C
Core 4 - 61.3C
CPU package - 67.8C

I have no doubt in my mind that the thermal issues with Ivy Bridge is due to the TIM route Intel decided to take with these CPUs, instead of soldering them like they used to do. If you really want to get the most of these processors and overclock them without burning the house down, you definitely need to look into de-lidding them and replacing the TIM with a high quality one. Obviously, I will recommend Coollaboratory's Liquid Metal Ultra based on MY findings. The past few weeks I have been trying like a madman, to get a stable overclock of 4.9GHz, but the temps were through the roof in the upper 90s on full load with the same 1.40 voltage.

Dude, are you being serious right now?

 

[Idontcare]

That is hardened steel. Its your choice of course, but copper is much more thermally conductive, much softer and less likely to increase risk of cracking the die. Could you cancel the order and go with copper instead? Just a suggestion. You never see steel being used for the base of heat sinks for the poor thermal conductivity. Steel really holds onto the heat. It might act as an insulator.

I think you misunderstand what the metal shim is for, understandable because the posts have been going up in rapid succession.

The shim is to lift the IHS off of the PCB such that the resulting gap between the underside of the IHS and the CPU silicon die is the same gap hieght as existed when the original stock CPU TIM was in place.

^ right now I have the situation on the right, I want the situation on the left, I aim to accomplish that by shimming the IHS off of the PCB.

To do that my shim must first of all be thick enough to fill the current gap between the IHS and the PCB - that gap is 0.14mm.

I need the shim to further lift the IHS another 0.06mm so that the gap between the IHS and the CPU is then 0.06mm, the same as was present when the original TIM was present.

Then and only then can I test a CPU TIM replacement, such as NT-H1, and generate temperature data that is apples-to-apples comparable to the temperature generated with the stock Intel CPU TIM.

At this point in time, for all I know the 20°C improvement I observed is because I decreased the gap between the IHS and the CPU by 0.06mm versus the improvement coming from the replacement of the original CPU TIM.

We all suspect (assume) the TIM is the culprit, but it may just be that it is the gap itself that is the culprit for the elevated temps with a stock IB...and naturally we are right to assume the answer will be that it is partly the fault of both.

It seems implausible that Intel went to all the expense of using a "boutique" TIM like NT-H1 or MX-4 given that they don't bother to do it for their stock HSF TIM either, so it seems plausible to assume the stock CPU TIM is inferior to boutique TIMs.

At the same time a gap of 0.06mm is huge, that is a lot of TIM to shuffle the heat through. Reducing that gap by 50% or 75% will bring large improvements to the thermal conductivity of the stack, that's just basic heat transfer mechanics there.

So I want to shim the IHS, and I don't want a ductile metal like copper which will compress and bend under applied downforce. I wanted a high-carbon steel that would stand the best chance of retaining its physical dimensionality when compressed.

I think it will become a lot more clear what I am attempting to accomplish with the shim once I have the material in hand and can post some pics of the shim in place.

 

[HondaCop]

Dude, are you being serious right now?

Why you say that?

- Grammar *censored*, this was sent from my phone.

 

[BonzaiDuck]

LinX test (21000 | 20 runs) with a voltage of 1.40v @ 4.9GHz:

Core 1 - 60.6C
Core 2 - 63.4C
Core 3 - 68.2C
Core 4 - 61.3C
CPU package - 67.8C

I have no doubt in my mind that the thermal issues with Ivy Bridge is due to the TIM route Intel decided to take with these CPUs, instead of soldering them like they used to do. If you really want to get the most of these processors and overclock them without burning the house down, you definitely need to look into de-lidding them and replacing the TIM with a high quality one. Obviously, I will recommend Coollaboratory's Liquid Metal Ultra based on MY findings. The past few weeks I have been trying like a madman, to get a stable overclock of 4.9GHz, but the temps were through the roof in the upper 90s on full load with the same 1.40 voltage.

Serious? IDC was able to get through some LinX runs with 1.37+V, and we said that's like putting 1.44V through the Sandy Bridge.

But really, this is good, extra data. 4.8 Ghz with 1.31V is a lot closer to comfort. And it's a lot closer to the 1.29+ which IDC showed for 4.7 Ghz. So I'm wondering (did I miss reading a post or something?) what your Liquid Ultra gives us at 4.7 Ghz. Does it trim something off the voltage as well as the tempertures? Or do we get the same voltage with lower temperatures?

EDIT: I think you were showing a 3 millivolt drop in voltage for 4.8 with Liquid Ultra as opposed to a so-so-TIM or the original TIM?

 

[Diogenes2]

Why you say that?
.

Those temps are really unbelievably low for the voltage and speed you are reporting..

 

[oceanside]

LinX test (21000 | 20 runs) with a voltage of 1.40v @ 4.9GHz:

Core 1 - 60.6C
Core 2 - 63.4C
Core 3 - 68.2C
Core 4 - 61.3C
CPU package - 67.8C

I have no doubt in my mind that the thermal issues with Ivy Bridge is due to the TIM route Intel decided to take with these CPUs, instead of soldering them like they used to do. If you really want to get the most of these processors and overclock them without burning the house down, you definitely need to look into de-lidding them and replacing the TIM with a high quality one. Obviously, I will recommend Coollaboratory's Liquid Metal Ultra based on MY findings. The past few weeks I have been trying like a madman, to get a stable overclock of 4.9GHz, but the temps were through the roof in the upper 90s on full load with the same 1.40 voltage.

I find your numbers very hard to believe. I have slightly higher temps than that on LinX fully loaded with a Sandy 2500k (soldered lid and all) @ 4.4GHz & 1.26vcore on big air. Noctua DH-14 (on high) and a Silverstone FT02B. Ambient doesn't go much past 25C here w/o AC.

If you have a H100, that is very much comparable to a DH-14.

What software are you reading the temps from?

 

[BonzaiDuck]

Yeah, I didn't look closely enough at his temperatures, but is it outside the realm of possibility? With a so-so TIM, IDC was showing (if I recall . . ) about 74 to 75C for the hottest core at 4.7.

You'd figure the Liquid Pro would show lower temperatures at 4.7 and voltages around 1.30. You just wouldn't figure to have the expected temperatures at 4.9 and 1.37 . . .

The other thing I wonder about, since he used the same metal TIM in its traditional application, is whether he lapped off the nickel plate on the IHS and lapped his water-block flat.

 

[cbn]

~500 micron.

The analogy is this - imagine the silicon substrate is a 10 story tall building. The active circuitry would be a 6" layer sitting at the very top of that 10-story building.

Then they flip it upside down and solder it onto the PCB.

The stuff we are doing on the backside of the silicon substrate has virtually no ability to effect the active circuits beyond that of creating cracks (inducing mechanical failure) or creating a thermal/electrical gradient that becomes problematic to first or second-order.

IDC, If we look at the following picture of a "Ready transistor" from "How computer chips are made"--> http://apcmag.com/picture-gallery-how-a-chip-is-made.htm?page=5

The 500 micron distance you gave would be equivalent to the dark gray portion directly under "green doped area"?

Then the metal layers from this step are added ---> http://apcmag.com/picture-gallery-how-a-chip-is-made.htm?page=7

.....And the chip is flipped over to be attached to the PCB. (With that 500 micron thick dark gray part facing up ) ??

Is this understanding correct??

 

[Markfw]

I got a couple of "factory unlidded" (they didn't have one) Xeon Nejalem CPU's and the only way I could get a HSF that worked, was by buying some very thing copper sheeting, and putting two layers over the core and under the HSF.

 

[moonbogg]

I think you misunderstand what the metal shim is for, understandable because the posts have been going up in rapid succession.

The shim is to lift the IHS off of the PCB such that the resulting gap between the underside of the IHS and the CPU silicon die is the same gap hieght as existed when the original stock CPU TIM was in place.

^ right now I have the situation on the right, I want the situation on the left, I aim to accomplish that by shimming the IHS off of the PCB.

To do that my shim must first of all be thick enough to fill the current gap between the IHS and the PCB - that gap is 0.14mm.

I need the shim to further lift the IHS another 0.06mm so that the gap between the IHS and the CPU is then 0.06mm, the same as was present when the original TIM was present.

Then and only then can I test a CPU TIM replacement, such as NT-H1, and generate temperature data that is apples-to-apples comparable to the temperature generated with the stock Intel CPU TIM.

At this point in time, for all I know the 20°C improvement I observed is because I decreased the gap between the IHS and the CPU by 0.06mm versus the improvement coming from the replacement of the original CPU TIM.

We all suspect (assume) the TIM is the culprit, but it may just be that it is the gap itself that is the culprit for the elevated temps with a stock IB...and naturally we are right to assume the answer will be that it is partly the fault of both.

It seems implausible that Intel went to all the expense of using a "boutique" TIM like NT-H1 or MX-4 given that they don't bother to do it for their stock HSF TIM either, so it seems plausible to assume the stock CPU TIM is inferior to boutique TIMs.

At the same time a gap of 0.06mm is huge, that is a lot of TIM to shuffle the heat through. Reducing that gap by 50% or 75% will bring large improvements to the thermal conductivity of the stack, that's just basic heat transfer mechanics there.

So I want to shim the IHS, and I don't want a ductile metal like copper which will compress and bend under applied downforce. I wanted a high-carbon steel that would stand the best chance of retaining its physical dimensionality when compressed.

I think it will become a lot more clear what I am attempting to accomplish with the shim once I have the material in hand and can post some pics of the shim in place.

Yup, I did't get it. Sorry about that. But what do you think of the idea anyway of using copper ontop of the die between the heat spreader? Maybe copper would be a good replacement for solder?

 

[Idontcare]

IDC, If we look at the following picture of a "Ready transistor" from "How computer chips are made"--> http://apcmag.com/picture-gallery-how-a-chip-is-made.htm?page=5

The 500 micron distance you gave would be equivalent to the dark gray portion directly under "green doped area"?

Then the metal layers from this step are added ---> http://apcmag.com/picture-gallery-how-a-chip-is-made.htm?page=7

.....And the chip is flipped over to be attached to the PCB. (With that 500 micron thick dark gray part facing up ) ??

Is this understanding correct??

Yep, you got it. And the total height of the circuitry including the wires might be all of 10 microns at the most.

One minor detail to point out, the dark gray area is actually ~800um when the wafer is in the fab and the circuits are being built up. Once all the circuits are finished then the wafer will eventually go through what is called a backgrind process to thin the wafer prior to dicing and packaging.

For IB, apparently Intel thins the wafers by about 250um, thinning the "dark gray" area in your pics to ~500micron.

Yup, I did't get it. Sorry about that. But what do you think of the idea anyway of using copper ontop of the die between the heat spreader? Maybe copper would be a good replacement for solder?

That's a good question. Provided the copper shim is smooth and the IHS is smooth such that you don't need to put TIM between those interfaces then I'd expect the performance to be quite comparable to that of these liquid metal TIMs.

 

[Idontcare]

This thread is on Anandtech main page via the twitter feed because it is awesome

(just when you think twitter is absolutely useless!)

wow! I'm honored, and impressed. Thanks blckgrffn for bringing it to Anand's attention, and thanks Anand for the retweet :thumbsup:

 

[exar333]

wow! I'm honored, and impressed. Thanks blckgrffn for bringing it to Anand's attention, and thanks Anand for the retweet :thumbsup:

This is the best thread I have read in over a year. Thanks IDC.

After reading this, I almost went out and just ordered a IB just try all the cool stuff you walked through in your process, just for the fun of it!

 

[Rifter]

You need to find a way to make that spinning IHS your avatar pic.

 

[BonzaiDuck]

Yep, you got it. And the total height of the circuitry including the wires might be all of 10 microns at the most.

One minor detail to point out, the dark gray area is actually ~800um when the wafer is in the fab and the circuits are being built up. Once all the circuits are finished then the wafer will eventually go through what is called a backgrind process to thin the wafer prior to dicing and packaging.

For IB, apparently Intel thins the wafers by about 250um, thinning the "dark gray" area in your pics to ~500micron.



That's a good question. Provided the copper shim is smooth and the IHS is smooth such that you don't need to put TIM between those interfaces then I'd expect the performance to be quite comparable to that of these liquid metal TIMs.

Put some pennies on the railroad track and then attempt to lap the best ones following the afternoon progression of container boxes?

 

[HondaCop]

For those having doubts about the numbers I posted last night, let me remind you that I am de-lidded and both the top of the IHS and the bottom of the H100's water block have been lapped to almost a mirror like shine. Liquid Metal Ultra was used on both the die to IHS connection and IHS to water block connection. Also, I used AIDA64 for the temperature readings. AIDA64 keeps track of max/low readings and gives you a real time average temp reading.

As soon as I start Linx, I reset all temps to delete previous readings.

I will continue testing late tonight because I've got court this morning and extra work all day after 12pm.

Oh another thing, right before going to bed last night, I did a quick run of Prime95's torture test, the one that stresses the cpu to the max, where prior to LMU I would be getting temps in the 90s within 10 minutes of the test at 4.8Ghz. Well, I ran it for ten minutes at 4.9Ghz, 1.40v and in ten minutes, temps were in the 70s! Never broke 80. But I had to sleep and shut it off.

- Grammar *censored*, this was sent from my phone.

 

[HondaCop]

Finally back home from court but I have to leave at noon again, Dammit! lol Yes, I am a cop but no, we don't drive Hondas. LOL Anyway, I turned the PC on and so far, Prime95 has been error free for the past 40 minutes while running the large, in-place torture test with the following settings:

4.9Ghz @ an average of 1.43v. I say average because I am overclocking with offsets and not fixed voltages. Also, I have all power options on the cpu enabled, so on idle, my cpu falls back to 1.6Ghz and very low voltages.

Here are the numbers so far in this current 4.9Ghz run (51 minutes and counting):

Temperatures

Item Min Max Average
CPU1 29 69 62.8
CPU2 21 73 65.9
CPU3 27 78 71.1
CPU4 27 71 63.7
CPUpkg 30 79 71.2


Voltages

Item Min Max Average
CPUVID 1.166 1.271 1.265
CPUcore 1.024 1.448 1.430
CPUPLL 1.587 1.600 1.597
DIMM 1.325 1.337 1.336
VCCIO 1.038 1.075 1.053

As you can see, even at an average vcore of 1.430v and running at 4.9Ghz, the highest any of the cores have reached is 78C and that was core 3 which is usually the hottest running one. The highest core 1 has reached is 69C!

Another thing that I've noticed is that for the first time, I can hear the fans on the H100 SPEEDING UP AND DOWN as the load changes! Never did I hear them do that ever since I built this PC a couple weeks ago. That tells me that now the heat is being transferred entirely to the IHS whereas before, most of the heat was staying on the die. This Liquid Metal Ultra is truly the shihznits. It's the next best thing to actually soldering the IHS to the cpu die.

 

[Phynaz]

Honda, how could your CPU package be hotter than any cores?

 

[HondaCop]

Honda, how could your CPU package be hotter than any cores?

Right now, the average temp for CPU Package stands at 71.4C. The highest average temp for the hottest core, which is core 3, stands at 71.3. I guess the CPU package will be as hot as the hottest core. which makes sense. As for the package being .1C hotter, you have to give the software some margin of error.

Here is a pic of my screen (sorry for the cell pic, but this is a gaming PC and I don't have Paint or Photoshop on it):

 

[Denithor]

Put some pennies on the railroad track and then attempt to lap the best ones following the afternoon progression of container boxes?

Only if you used some pretty old pennies. The ones today are mostly zinc...



http://www.usmint.gov/about_the_mint/fun_facts/?action=fun_facts2

 

[AnonymouseUser]

Here is a pic of my screen (sorry for the cell pic, but this is a gaming PC and I don't have Paint or Photoshop on it):

You should have the Snipping Tool installed by default.

 

[HondaCop]

You should have the Snipping Tool installed by default.

Oh I forgot about that one! I need to check because I also remove a bunch of default stuff from my Windows installation as part of my optimization process.

- Grammar *censored*, this was sent from my phone.