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

[richierich1212]

Awesome guide and thanks for the highly detailed thread

 

[HondaCop]

Here's one that IDC did on lapping his H100. link

Thanks!

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

 

[ShadowVVL]

Nice work IDC, this is a vary cool thread kind of makes me wanna go get a 3770k and delid it.
BTW
You can see kinda what IDC looks like in the beginning of this video in full screen @720 on the mirrored lid.Is he fully dressed in a fab uniform?



http://www.youtube.com/watch?v=Eeq2_TDxZlA

 

[BonzaiDuck]

It at least proves that you can OC to 4.7 at a fairly low voltage, or below 1.30V with temperatures in the mid-'70s C. And he's using a so-so TIM replacement.

So the question remains: If you could trim another 10C from the load temperatures, what would the voltage be for either the same over-clock setting or higher -- like 4.8Ghz?

The drop in temperatures @ 4.7 Ghz meant a 0.04V drop -- not trivial.

Overall, if you're going to build a computer this year with plans to overclock the processor, you would buy the Ivy Bridge and it wouldn't make sense if you did not de-lid it and use a better thermal compound. Unless, of course, the extra work seems too much . . .

And if the extra work seems too much, you wouldn't be that serious about over-clocking.

 

[pantsaregood]

What kind of voltage did you need for 4.8 GHz? 4.7 GHz at 1.256v seems pretty reasonable, and that's a little beyond what most Sandy Bridge cores will hit on reasonable voltage. Above 4.5 GHz, my i5-2500K needs pretty major voltage jumps to remain stable. 4.4 GHz will run stable at 1.3v (~1.36v before Vdroop), but 4.5 GHz requires ~1.34v after Vdroop. Above that, I have to jump to >1.4v, which I'm not comfortable with.

You may be showing us that Ivy Bridge scales slightly better than Sandy Bridge. Sandy Bridge usually makes it to 4.4-4.6 GHz relatively easily, but 4.7-4.8 GHz usually require high voltage. 4.9-5.0 GHz is probably attainable on a large portion of Sandy Bridge cores, but the voltage required is far out of the "safe" range. If Ivy Bridge manages 4.7-4.9 GHz pretty consistently, then no complaints of Intel's 22nm process being a failure are valid.

 

[Denithor]

Amazing.

Just amazing!

So much for all the people posting saying that IB higher temps were due to the smaller node concentrating the heat and not allowing it to dissipate effectively. Morons.

Next up, direct core cooling and then one of the liquid metal TIMs. Cannot wait.

Awesome work IDC!!!

 

[BonzaiDuck]

What kind of voltage did you need for 4.8 GHz? 4.7 GHz at 1.256v seems pretty reasonable, and that's a little beyond what most Sandy Bridge cores will hit on reasonable voltage. Above 4.5 GHz, my i5-2500K needs pretty major voltage jumps to remain stable. 4.4 GHz will run stable at 1.3v (~1.36v before Vdroop), but 4.5 GHz requires ~1.34v after Vdroop. Above that, I have to jump to >1.4v, which I'm not comfortable with.

You may be showing us that Ivy Bridge scales slightly better than Sandy Bridge. Sandy Bridge usually makes it to 4.4-4.6 GHz relatively easily, but 4.7-4.8 GHz usually require high voltage. 4.9-5.0 GHz is probably attainable on a large portion of Sandy Bridge cores, but the voltage required is far out of the "safe" range. If Ivy Bridge manages 4.7-4.9 GHz pretty consistently, then no complaints of Intel's 22nm process being a failure are valid.

If his numbers prove a positive interaction between temperature reduction and voltage required over something like a 94C -> 74C range, then you have to wonder what another 10C to 15C might give in that regard. One of the enthusiast-forum posts I cited, a Liquid-Pro application gave 69C, although I forget at the moment what his over-clock speed setting and voltage had been. I'm also guessing there are better water-cooling choices than the Corsair H100, only for providing a few more degrees reduction.

That leaves the "direct-contact" option of removing the top panel of the IHS through sanding, or just eliminating the IHS altogether. And if you use a Liquid-Pro application that way, kiss off saving the processor if you want to remove the heatsink at some point. Maybe you could achieve almost the same improvement using micronized diamond between the IHS and die, and Liquid-Pro or Indigo Xtreme between the HSF and IHS.

To achieve what seems like a guaranteed advantage, it means lapping the IHS and heatsink base, removing the IHS and replacing the TIM, and using the same or better TIM between the base and IHS. After that, it becomes a choice between "air and water."

 

[Kenmitch]

So I pulled off the H100 and unmounted the IHS to move on with the next test - using the Intel CPU TIM from under the IHS as a TIM on top of the IHS of my 2600k for a comparison with other TIMs. (i.e. the "is that CPU TIM crappy stuff, or good stuff?" test)

I'm also interested in this part of the testing. Did you post results yet or are you holding out on them? Wondering if it's a combo of the TIM being not so good and possibly the bonding agent for the heatspreader increasing the gap too much.

Interesting results so far

 

[pantsaregood]

Amazing.

Just amazing!

So much for all the people posting saying that IB higher temps were due to the smaller node concentrating the heat and not allowing it to dissipate effectively. Morons.

Next up, direct core cooling and then one of the liquid metal TIMs. Cannot wait.

Awesome work IDC!!!

This doesn't make thermal density any less relevant. Even after he's done all of this, his CPU is running hotter than my 4.4 GHz 2500K on a Hyper 212+ with all my case fans on low. While LinPacking, my CPU never breaks 70C.

That's with a cheap air cooler and Arctic Silver 5 I bought in 2005 to put on an Athlon 64.

 

[BonzaiDuck]

This doesn't make thermal density any less relevant. Even after he's done all of this, his CPU is running hotter than my 4.4 GHz 2500K on a Hyper 212+ with all my case fans on low. While LinPacking, my CPU never breaks 70C.

That's with a cheap air cooler and Arctic Silver 5 I bought in 2005 to put on an Athlon 64.

But he's not running at 4.4. The crucial speed or the one of most interest so far is 4.7. My cores go to almost 74C at 4.6 (i7-2600K) under full load with NH-D14 and diamond paste. They're well over 80C at 4.7. And IDC isn't even testing "best" TIM yet, while those untested options promise further improvements of 5C or more -- my guess.

Keep in mind that -- yes, there's been a die-shrink to 22nm process, but the TDP has also come down from 95W to 77W. So thermal density is "important," but it is one of several factors.

 

[dma0991]

If a good aftermarket TIM under the IHS would make such a huge difference in temps, would that mean that Intel is using the cheap stuff? :hmm:

 

[Puppies04]

So much for all the people posting saying that IB higher temps were due to the smaller node concentrating the heat and not allowing it to dissipate effectively. Morons.

So understanding physics makes someone a moron in your books, lol.

 

[BonzaiDuck]

If a good aftermarket TIM under the IHS would make such a huge difference in temps, would that mean that Intel is using the cheap stuff? :hmm:

They're "usin' what they're usin'" just to keep the thermals low enough for intended, stock speeds and voltages.

I got slammed for suggesting that they might do something to correct the short-coming by choosing either a better TIM or revising their production process for this line of processors.

The truth is -- yes -- the web is "aflame" with talk about the IB TIM and heatspreader. But it is aflame with remarks coming from a small part of the overall user-base. People who buy OEM computers aren't likely interested in the over-clocked potential of the processor, and OEMs are only going to be interested insofar as they can cheaply roll off mildly over-clocked gamer systems and a lot more systems running the CPU at stock specs.

We don't know whether Intel considered diamond paste, whether there is some "risk" of using micronized diamond (not too likely), whether they looked at the other exotic "metal" TIMs. IF it adds to their costs, choosing a less costly solution is their prerogative, and makes business sense. It makes business sense if the majority of processors you sell are going to OEM makers.

UPDATE: I'd bet a few bucks that -- as I insinuated under "withering criticism" of my theories of what Intel might "do" before the next CPU-generation -- they "walk among us." "They Live." "The Threads have Eyes." They might even pay someone to gather Intel's "intel" about Intel from Intel-enthusiast web-sites. Their engineers may include an "over-clocker" component, so another hole in the same part of "sieve of probabilities."

So you wonder what they might do to support the Haswell cores, at even less TDP. They could do "more," or they could do "less." They could make us "do more or less." Maybe they might think they can afford a $5 increase in unit production cost, especially if they figure out a way to reduce it elsewhere. How many are they going to sell, and what profit margin can they expect?

 

[richaron]

We don't know whether Intel considered diamond paste, whether there is some "risk" of using micronized diamond (not too likely), whether they looked at the other exotic "metal" TIMs. IF it adds to their costs, choosing a less costly solution is their prerogative, and makes business sense. It makes business sense if the majority of processors you sell are going to OEM makers.

It makes sense if they already hold the high ground with performance & efficiency, while IB is an incremental upgrade. They will sell plenty no matter what.

They are/were already eating into their own profits with ~mid range~ chips which can be easily be clocked to ludicrous speed.

 

[ShintaiDK]

If a good aftermarket TIM under the IHS would make such a huge difference in temps, would that mean that Intel is using the cheap stuff? :hmm:

Remember while removing the adhesive. He also lowered the height of the IHS. So it got closer to the die and gives a better effect.

Plus we got to remember that Intels solution needs to work for 25 years+, it needs to work everywhere on the earth and basicly under any situation.

 

[ALIVE]

Amazing.

Just amazing!

So much for all the people posting saying that IB higher temps were due to the smaller node concentrating the heat and not allowing it to dissipate effectively. Morons.

Next up, direct core cooling and then one of the liquid metal TIMs. Cannot wait.

Awesome work IDC!!!

well i read some post in this forum and there was a person asking the same question okey the area in smaller but also the comsumption so that balance more or less each other. and they accused him nope you do not know what you say.

i think it is beyond any doubt any longer
ivy is hotter cause intel squeezed a peny or intel wanted the people not able to overclock ivy that much.

case closed
and some apologies much be given to some people attacked so furiously in these forums

 

[ALIVE]

But he's not running at 4.4. The crucial speed or the one of most interest so far is 4.7. My cores go to almost 74C at 4.6 (i7-2600K) under full load with NH-D14 and diamond paste. They're well over 80C at 4.7. And IDC isn't even testing "best" TIM yet, while those untested options promise further improvements of 5C or more -- my guess.

Keep in mind that -- yes, there's been a die-shrink to 22nm process, but the TDP has also come down from 95W to 77W. So thermal density is "important," but it is one of several factors.

thermal density increased close to 10% temped skyrocket even more

 

[BonzaiDuck]

thermal density increased close to 10% temped skyrocket even more

So? If they released the K Ivy Bridgers with this process -- they can improve it in the next round, even with the higher thermal density and TDP lower than Ivy. Same or faster speeds with lower voltage.

You know -- the other thing about the business end of this -- if the majority of processors sold go to buyers who don't care about OC'ing, the processors will use even less power. The power savings may be more important in the marketplace.

 

[Puppies04]

well i read some post in this forum and there was a person asking the same question okey the area in smaller but also the comsumption so that balance more or less each other. and they accused him nope you do not know what you say.

No they don't, the size decrease was a lot higher than the power consution savings.

i think it is beyond any doubt any longer ivy is hotter cause intel squeezed a peny or intel wanted the people not able to overclock ivy that much.

No it isn't beyond any doubt,you just took your small amount of knowledge on a subject and decided you understood it, you were wrong. Also you seem to lack the basic understanding of the difference between heat and temperature, go read a book. As for your ranting about this being a conspiracy by intel to prevent IB being OC'd too high you need to realise that intel already explained why they changed the TIM and I have yet to see one person who understands the issue prove that they were lying.

case closed

I think not

and some apologies much be given to some people attacked so furiously in these forums

Lol

 

[pantsaregood]

But he's not running at 4.4. The crucial speed or the one of most interest so far is 4.7. My cores go to almost 74C at 4.6 (i7-2600K) under full load with NH-D14 and diamond paste. They're well over 80C at 4.7. And IDC isn't even testing "best" TIM yet, while those untested options promise further improvements of 5C or more -- my guess.

Keep in mind that -- yes, there's been a die-shrink to 22nm process, but the TDP has also come down from 95W to 77W. So thermal density is "important," but it is one of several factors.

At 4.7 GHz, I have to push my voltages to the point I hit around 78 C. My CPU certainly isn't blowing up, but the voltages required are higher than I'd like. This is on a Hyper 212+, however. An H100 would run significantly cooler.

All IDC has done is show that poor thermal compound is an issue on his unit. There have been reports of some Ivy Bridge CPUs responding relatively poorly to a change in TIM. This definitely shows what a poor paste job can do, but it doesn't make thermal density any less of an issue.

That said, this is a perfect example of what could potentially happen when you go beyond what the manufacturer intends. A 20 C drop in temperature and bypassing previous thermal limitations on your CPU - definitely nice.

Also, to all conspiracy theorists: Ivy Bridge is subject to increased thermal density. It isn't a conspiracy. The lack of consistency in temperature drops from applying new TIM to the IHS further disproves this ridiculous conspiracy. If anything, all you're showing is that Intel's manufacturing process isn't great, not that they're conspiring behind your back.

 

[BonzaiDuck]

At 4.7 GHz, I have to push my voltages to the point I hit around 78 C. My CPU certainly isn't blowing up, but the voltages required are higher than I'd like. This is on a Hyper 212+, however. An H100 would run significantly cooler.

All IDC has done is show that poor thermal compound is an issue on his unit. There have been reports of some Ivy Bridge CPUs responding relatively poorly to a change in TIM. This definitely shows what a poor paste job can do, but it doesn't make thermal density any less of an issue.

That said, this is a perfect example of what could potentially happen when you go beyond what the manufacturer intends. A 20 C drop in temperature and bypassing previous thermal limitations on your CPU - definitely nice.

Also, to all conspiracy theorists: Ivy Bridge is subject to increased thermal density. It isn't a conspiracy. The lack of consistency in temperature drops from applying new TIM to the IHS further disproves this ridiculous conspiracy. If anything, all you're showing is that Intel's manufacturing process isn't great, not that they're conspiring behind your back.

I wouldn't call the speculations about "conspiracy:" It's a duopoly, and the two major producers can game against each other. But Intel probably has minor interest in what the over-clocking segment does, except for "gathering intelligence" about their processors.

They should certainly know by now, however, that they can deal somewhat or partly with any Haswell thermal density by correcting the TIM problem. What are they doing? Picking an "adequate" TIM, assuring the product runs within spec -- not much more than that.

 

[Idontcare]

Gosh Thanks everyone for all your kind words! :$

When you put the IHS back on your chip, was there any gap at all between the lid and the PCB? I'm concerned that the IHS might rock and crush a corner or edge because there isn't black goop filling that space and preventing it from rocking.

I took lots of measurements with a micrometer to address this question and concern, working up a post on that right now. Should be up this evening before it gets too late. Haven't processed all the numbers yet so I have no quick answer for you right now, sorry

What was the stable load voltage (anything that survived a prelim IBT or LinX test) @ 4.9 with the i7-2600K? Wasn't the successive voltage increase for i7-2600K between 4.7 and 4.9 Ghz rising steeply?

The i7-2600K apparently overclocks to 4.9Ghz with approximate load voltage in the 1.44 to 1.46V range. In any event, it is above 1.41V.

So if IDC OC'd an i7-3770K chip at 4.9 with about 1.37+V, it doesn't seem like much of an advantage. . . .

At 4.9GHz, using the same mobo, H100, etc etc, my 2600k requires 1.444V for LinX stability, the tangible benefit appears to be the near 45W reduction in power-consumption when OC'ed to those extremes.

But I don't want to overstate the relevance of the power savings...at the ~$0.10/kWHr rate that I pay here in PA, if I ran my 3770k at 4.9GHz 24hrs a day, every day for an entire year then I'd be looking at saving around $37 for the entire year versus having used my 2600k in the exact same capacity.

I'll take it because I've already bought the chip, but $40/yr is not a decision-making datum point, but it is about the only "plus" to be had in the pro-IB column versus the pro-SB column when it comes to high-end computing performance at high clocks and high operating voltages. (IB wins hands-down when it comes to the other end of the power-usage spectrum, mobile platforms will definitely benefit from IB)

from these photos we can clearly make out that IDC runs his own fab facility in his basement. He is currently manufacturing 14nm chips and is developing a 10nm in his spare time for his own personal use using EUV lithography (also developed in his spare time)

in this photo we can see him working on a prototype OLED Ipad4 display. We see he is of asian descent, clearly only the finest genetic material is part of IDC's genome.

LOL, I do have a fuzzy/blurry photo of me in the fab, the following was a pic I took of me and Brian (a friend/co-worker of mine) of our reflection in the side of a wet cleans tool that was shrouded in mirror-polished stainless steel (I'm the dude on the left as you see the photo):

to play devils advocate, having /\/\/\/\/\/\/\/\ as in with an unlapped surface vs __________ with lapped would increase the surface area, allowing for greater thermal dissipation....if you could get the thermal compound into those grooves...no?

I think the answer to this is self-evident if you'll consider the following - if you filled those grooves with TIM, what would be the best TIM to use?

What if you used copper instead of TIM to fill in the grooves? Would heat transfer through the TIM better than through solid copper if those grooves were instead filled with solid copper?

The ridges may seem like a good idea for heat transfer until you take into account what it is that you are trying to transfer the from and into.

If the goal was to transfer the heat into the TIM because the TIM was superior for heat transfer versus the IHS that holds the heat then you would want the "VVVVV" surface. This is true for thermal transfer inside your water-block or for the fins of the HSF, you want to maximize the surface area for heat transfer from the copper block into the water or the aluminum fins to the air...but that is ONLY because once the heat transfers to the water/air then you physically move the water or air, you don't rely on the water or air itself to thermally conduct the heat away from the copper block.

But in a traditional IHS:TIM:HSF stack you are relying on the TIM to thermally conduct the heat away from the IHS, and replacing copper with TIM is not good unless the TIM itself has higher thermal conductivity than the copper (which is possible, but expensive, and would require using pure diamond or silver - and not a colloidal admixture containing bits of them, it would have to be solid material for phonon-phonon coupling reasons).

 

[pantsaregood]

A ridged surface (/\/\/\/\) doesn't work well because the effective surface contact is actually diminished, even though the surface area is greater.

If you had a heatsink with a base that was actually cut with ridges to fit inside of the IHS ridges, then more surface area would be great.

 

[Idontcare]

This is quite amazing. Thanks for all the hard work!

What I don't understand is how that razor blade didn't hit the die, which I assume is not flush with the PCB, when cutting the stock TIM.

BTW, do you have a third hand? No other explanation for how you got those pics!

Pretty sure it did hit the die actually, if you watch this short video clip the noise you hear when the blade slips and then comes to an abrupt stop is the blade stopping because it ran into the corner of the silicon die.

The fact that it did not result in a dead die, and no notable chipping or marring occured on the die from the collision with the blade, just goes to show how tough the silicon die really can be.

I wouldn't want to do it a second time though I've had silicon wafers break in my hands, I know how brittle that material really is, and I was quite surprised my little slip-up with the blade there did not kill the CPU, guess I'll count my lucky stars

Oh and the pics are easy to explain - tripod! Although at the rate this "sport" is evolving, I can foresee a need to get a go-pro headcam when Haswell comes out

Sticky. But you must finish this.

I'm surprised you got so much done while getting so much documented, videoed and photographed.

Do you have any indication that perhaps the Intel TIM is really AS5?

I've got more to put up yet, the documentation efforts are a bit tedious, not going to play that down here, but I enjoy it so its not bothersome or anything like that. What really takes the time though is condensing all the documentation data and test data into digestible nuggets of pic-goodness. A picture is worth a thousand words, I'm a visual learner myself, so I value getting the right photo or the graph just right so others can more easily absorb whatever this thread has to offer :thumbsup:

Regarding the CPU TIM (as I have come to call it) - pretty sure it is NOT AS5. You'll see what I mean later on when I get the post up, but it is very stiff/rigid material which is neither pliable nor malleable. It probably was at one point in time and then was "cured" so that it formed a more permanent structure of sorts, but AS5 it is not.

I also tried a carbon based pastes, mx4 and NT-H1. Both of these gave inferior results compared to AS5. You have any to try?

I do and I will. I have AS5, Ceramique, TX-2, MX-1, MX-4, NT-H1, and Indigo Xtreme. Not too mention my assortments of unconventional TIMs :sneaky:

Is there a reason you haven't tried direct die contact? I guess it isn't irrational to worry about cracking the die, but we've all mounted a Coppermine or K7 back in the day. I don't see why it would be any more dangerous now.

No reason, I intend to make an attempt at it, I had a positive experience in doing that with my GTX460 when I delidded it last year. The trick then is the same as it would be here, dealing with adjusting the HSF standoffs from the socket to account for the thickness of the IHS.

Adjusting the standoffs on my video card was pretty straightforward, cutting down the standoffs for my CPU's cooler is going to be much more challenging (and irreversible).

I'm also interested in this part of the testing. Did you post results yet or are you holding out on them? Wondering if it's a combo of the TIM being not so good and possibly the bonding agent for the heatspreader increasing the gap too much.

Interesting results so far

Yeah I would say both are in play - the question we'll never be able to answer is whether this situation is created by Intel for purposes of product longevity or simply for purposes of cost-reduction.

If it is only for financial reasons then us enthusiasts undoing what they did and redoing it ourselves as done here shouldn't be a problem for us...but if we are undoing something that was put in place for internally known reasons like thermal-cycling stress issues then we may be creating our very own "bumpgate" saga while thinking we are doing something neat.

Time will tell us whether or not the delidder-clan is full of folly or prudence.

What kind of voltage did you need for 4.8 GHz? 4.7 GHz at 1.256v seems pretty reasonable, and that's a little beyond what most Sandy Bridge cores will hit on reasonable voltage. Above 4.5 GHz, my i5-2500K needs pretty major voltage jumps to remain stable. 4.4 GHz will run stable at 1.3v (~1.36v before Vdroop), but 4.5 GHz requires ~1.34v after Vdroop. Above that, I have to jump to >1.4v, which I'm not comfortable with.

You may be showing us that Ivy Bridge scales slightly better than Sandy Bridge. Sandy Bridge usually makes it to 4.4-4.6 GHz relatively easily, but 4.7-4.8 GHz usually require high voltage. 4.9-5.0 GHz is probably attainable on a large portion of Sandy Bridge cores, but the voltage required is far out of the "safe" range. If Ivy Bridge manages 4.7-4.9 GHz pretty consistently, then no complaints of Intel's 22nm process being a failure are valid.

4.8GHz requires 1.325V and results in a peak operating temp of 81°C

 

[BonzaiDuck]

. . .
At 4.9GHz, using the same mobo, H100, etc etc, my 2600k requires 1.444V for LinX stability, the tangible benefit appears to be the near 45W reduction in power-consumption when OC'ed to those extremes.

But I don't want to overstate the relevance of the power savings...at the ~$0.10/kWHr rate that I pay here in PA, if I ran my 3770k at 4.9GHz 24hrs a day, every day for an entire year then I'd be looking at saving around $37 for the entire year versus having used my 2600k in the exact same capacity.

I'll take it because I've already bought the chip, but $40/yr is not a decision-making datum point, but it is about the only "plus" to be had in the pro-IB column versus the pro-SB column when it comes to high-end computing performance at high clocks and high operating voltages. (IB wins hands-down when it comes to the other end of the power-usage spectrum, mobile platforms will definitely benefit from IB)
. . .

1.44-something was what I gathered by looking at some old OC forum posts. And I'd think that 1.376V is excessive for the Ivy Bridge, just as we might incline for the 1.44V number on Sandy. That may (or may not) leave open the question about how it scales between 4.6 and 4.8Ghz. Or -- how a better choice of TIM and an additional drop in temperature might further mean the same overclock setting at lower voltage, or higher at the same voltage. This may be splitting hairs and certainly might not provide a reason to dump a Sandy for an Ivy. It might be good to find out. Or . . . I might be curious to find out . . .

So as I may have already said, if one doesn't have the Sandy, then buy the Ivy. If you want to over-clock the Ivy, one has a choice to do the IHS/TIM mod, or not to do it. But if one were going to over-clock the Ivy seeking the best result, it would only be the serious thing to do.

A serious thing to do which has a cost in time, effort, a few dollars for this and that, and some minor risk for all-thumbs clumsiness.