High CPU temp on new custom loop - what am I missing?

[Stilez]

I've built a nice low-restriction CPU loop, and if the Aquacomputer flow meter is to be believed Ive got flow rates about 3.3 litre/min (~ 0.72 GPM GPM) which is slightly on the low side, and components chosen for thermal excellence, but the CPU temp is still a lot higher than I expect.

Here's the loop:

Reservoir + pump (Heatkiller tube+D5 PWM)
> flow meter
> temperature sensor
> radiator (HW Labs Black Ice SR-2 480 with Thermaltake fans)
> filter (Barrow)
> block (Heatkiller IV pure copper + Noctua's thermal compound)
> back to reservoir + pump

Tubing is 16/10 EPDM and fittings are all Barrow. Except for right angles within the radiator and block, I've tried to make all flows as direct as I can. There's no sign of gunk or blockage. It's all new except the rad which I cleaned using Mayhem's Blitz and ran clear and clean afterwards. The thermal compound is Noctua's usual, which is a good one.

My thinking was, it uses one of the top few blocks out there, single loop 16/10 with no hard corners (except internal to rad and block), copper everywhere, decent flow rate, good fans at max speed... yet on stress testing, I get CPU core + package temp 75 C in HWInfo, vs. water temp 30 C (ambient is 25 C).

What am I missing? Why are the CPU and water temp so hugely different? Or is that normal?

Is there anything I can do, to get significantly more heat out of the CPU and into the water?

 

[seth1969]

What cpu is it bud?

 

[wookiee87]

given that you have a asus x299 sage your rocking a lga 2066 cpu, depending on which model you have they will all pretty much run warm regardless of what cooling you have, the thermal paste used between the die and the ihs is ok but if you want the best cooling results a delid is needed.

i had a 7900x and running stock was in the mid 70's under a custom 360m rad and overclocked 4.4ghz all core quickly hit around 90 degrees, i took the decision to have my cpu delidded and when i got it back it had a copper ihs fitted and liquid metal applied and under a new 4.8ghz all core clock my temps on a 360mm custom loop were in the mid 70's.

 

[Stilez]

What cpu is it bud?

i9-10980xe. 18 core.
Right now it's stress testing, so it's OC to about 44 ~ 45x across all cores running non-AVX prime95. Seems pretty stable.

(Its also dumping an insane amount of heat; HWInfo reckons 327 W @ 88 C core/package. But water temp is just 30 C)

given that you have a asus x299 sage your rocking a lga 2066 cpu, depending on which model you have they will all pretty much run warm regardless of what cooling you have, the thermal paste used between the die and the ihs is ok but if you want the best cooling results a delid is needed.

i had a 7900x and running stock was in the mid 70's under a custom 360m rad and overclocked 4.4ghz all core quickly hit around 90 degrees, i took the decision to have my cpu delidded and when i got it back it had a copper ihs fitted and liquid metal applied and under a new 4.8ghz all core clock my temps on a 360mm custom loop were in the mid 70's.

This one's soldered so delid isn't possible.

So my question is really about if that huuuge gap between cpu temp and water temp is "normal"/reasonable/expected, of about 45 C? Like, I expected to pull more heat out of the CPU into the water, and have I missed a trick somehow that could further improve the CPU temp drop?

 

[wookiee87]

i9-10980xe. 18 core.
Right now it's stress testing, so it's OC to about 44 ~ 45x across all cores running non-AVX prime95. Seems pretty stable.

(Its also dumping an insane amount of heat; HWInfo reckons 327 W @ 88 C core/package. But water temp is just 30 C)


This one's soldered so delid isn't possible.

So my question is really about if that huuuge gap between cpu temp and water temp is "normal"/reasonable/expected, of about 45 C? Like, I expected to pull more heat out of the CPU into the water, and have I missed a trick somehow that could further improve the CPU temp drop?

are you able to increase the pump speed, likely the cpu is throwing out so much heat under load the loop isn't fast enough taking heat away and replacing with cool liquid.
tbh a overclocked 18 core cpu with hyperthreading will heat up very fast and no matter the cooling loop most will struggle, and if you can keep it around 75 degrees under load thats actually pretty good for such a high core count chip.

also worth noting your cpu block may be installed incorrectly, most x299 setups have the barbs side by side and usually the barb to the left is inlet and to the right is outlet, but the heatkiller block you have is slightly different try you block with inlet to the right and outlet to the left

install the block so the heatkiller writing is the correct way up like below

 

[seth1969]

i9-10980xe. 18 core.
Right now it's stress testing, so it's OC to about 44 ~ 45x across all cores running non-AVX prime95. Seems pretty stable.

(Its also dumping an insane amount of heat; HWInfo reckons 327 W @ 88 C core/package. But water temp is just 30 C)


This one's soldered so delid isn't possible.

So my question is really about if that huuuge gap between cpu temp and water temp is "normal"/reasonable/expected, of about 45 C? Like, I expected to pull more heat out of the CPU into the water, and have I missed a trick somehow that could further improve the CPU temp drop?

I would say its fine, a lot of heat in a small package and your block is as good as it gets.
As above try mounting the block goofy if still not happy whack another 480 rad on your board and up the pump speed.

 

[seth1969]

Is your m/b cpu in a case?

 

[Stilez]

are you able to increase the pump speed, likely the cpu is throwing out so much heat under load the loop isn't fast enough taking heat away and replacing with cool liquid.
tbh a overclocked 18 core cpu with hyperthreading will heat up very fast and no matter the cooling loop most will struggle, and if you can keep it around 75 degrees under load thats actually pretty good for such a high core count chip.

also worth noting your cpu block may be installed incorrectly, most x299 setups have the barbs side by side and usually the barb to the left is inlet and to the right is outlet, but the heatkiller block you have is slightly different try you block with inlet to the right and outlet to the left
install the block so the heatkiller writing is the correct way up like below

I agree that it's handling a lot of heat and there's a limit. But I'm not having luck keeping the temps as low as I'd expect even so. I'm getting high 80s (87-90 C) not under 75 C at all.

Granted its OC but still..... almost 60 C difference between reported package temp to water temp, when using one of the better blocks and compounds out there? That's kinda more than unexpected...


Pump speed:

I don't know if I can up the pump speed. 3.3 L/min (0.7 GPM) might be about right for a D5 on this simple loop?

Annoyingly the PWM and tacho were both dead on arrival (tried all baseboard connectors and wired PWM to +5v molex as a last resort which should be 100% duty cycle on PWM, none helped or changed speed up or down). But that said, 3.3 L/min (0.7 GPM) isnt totally unreasonable. Should a D5 could get much higher on this loop? It doesn't sound like it's idling at 40-60% as some pumps do. I could buy and try out a D5 vario for certainty, but does it sound likely?

Block orientation/connection:

Watercool's instruction docs for the block state that "The HEATKILLER® IV has no predefined orientation on the CPU. Depending on the mainboard layout, all four facings should be possible. Those are: logo left (vertical), logo right (vertical), logo on top (horizontal) and logo at the bottom (horizontal). The alignment of the cooler has almost no impact on cooling performance and may be chosen freely", and "The intake is in the center of the cooler." (sections 3+10)

That's what I've followed. I just rechecked to make sure I did get it right. Intake is indeed centre and right now the PC case is moved with its back flat on the floor so there's no "up" or "down", motherboard and CPU are parallel to the floor. So orientation shouldn't matter and should be eliminated as a factor regardless? What do you think?

I would say its fine, a lot of heat in a small package and your block is as good as it gets.
As above try mounting the block goofy if still not happy whack another 480 rad on your board and up the pump speed.

I'm not sure another rad or faster water flow will add that much? If it was having trouble dumping heat out of the loop, I'd expect to see higher loop temps compared to ambient.

But here I'm seeing water temps barely 6 degrees above ambient, but CPU temps a full 55-60 degrees above water temp. That suggests to me, that the issue is heat transfer at the cpu/block interfaces, not rad->air.

If another rad dropped the water delta T by say half, to say 27 C not 30C, or faster water flow dropped another couple of degrees in the water, the difference at the *block* end would be 58 C not 55 C - almost negligible, not enough to force much more heat transfer from the CPU.

Does that sound about right?

Is your m/b cpu in a case?

The board and CPU are in a case, but the case is *wide* open to ambient, and there are full-on fans blowing at the entire CPU/VR region. Full view:

Other/"exotic" options? (Speculative!)

I'm using Noctua NT-H1 at present ("thin spread" rather than "dot" due to the large size of the 2066 heat spreader: the die is central but every bit of contact helps?). One way to win a couple more degrees might be improving the thermal compound layer. Not really the bigger change I'm looking for though. This article suggests an "X" shape instead for a tiny improvement. Another option I've read about might be Coollaboratory Liquid Metal Pad, a kind of metal leaf that melts into a thermal layer when first heated and is said tomwin a couple of.degrees over the noctua. (I won't try a fluid metal compound, too inexperienced to judge quantity on this size CPU).

Another option might be to peltier cool the water loop a bit, perhaps via the rad body itself near the inlets/outlets, where there's a large metal contact area, rather than the block, to take a bit more heat out than the rad alone can do, and maybe slightly chill it a little below ambient. Too much risks condensation but EPDM tubing is a pretty good insulator which works both ways, and if peltier+rad drops the water loop from 6 above ambient to maybe 5-8 below, perhaps that differentual will pull noticeably more heat from the CPU without risk to the system? (Most of the loop is outside the case and "condensation safe", see pics).

I guess both of these would take me into exotic cooling territory. Not usual for a 24/7 workstation. But if the issue is heat transfer from CPU package to water, as it seems, and I'm using a good rad and one of the better blocks/TIMs, is there anywhere else to look for a further 5+ C gain, without it?

Can I achieve much by say, doubling pumps, flow rates, rad "push" fans, or something else I'm missing?

Is it usual for there to be such a huge temp difference between CPU package and water temps if I've done everything correctly with good quality components?

 

[wookiee87]

Block orientation/connection:

Watercool's instruction docs for the block state that "The HEATKILLER® IV has no predefined orientation on the CPU. Depending on the mainboard layout, all four facings should be possible. Those are: logo left (vertical), logo right (vertical), logo on top (horizontal) and logo at the bottom (horizontal). The alignment of the cooler has almost no impact on cooling performance and may be chosen freely", and "The intake is in the center of the cooler." (sections 3+10)

That's what I've followed. I just rechecked to make sure I did get it right. Intake is indeed centre and right now the PC case is moved with its back flat on the floor so there's no "up" or "down", motherboard and CPU are parallel to the floor. So orientation shouldn't matter and should be eliminated as a factor regardless? What do you think?

regarding the instructions for block installation the orientation of the block can have a massive impact on cpu temps, you need the base plate of the block where the fins are cut into the metal ideally parallel to the cpu die direction that way you maximize the cooling area of the block relevant to the cpu, to find that out you need to take the block apart and see which direction the fins are pointing.

if the base plate fins are in a vertical position that is the wrong way, as a small area will make contact with the ihs and directly under the die, you need the base plate fins to be in a horizontal position as that way the whole ihs is being cooled down by the block, you'd see better load temps.

if the above has you confused don't worry, the long and short of it watercool's manual isn't very clear on which way the block should install, i still recon that it's not installed correctly and rotating the block 90 degrees to the right will likely help (take the block off and turn it 90 degrees to the right and re install it, don't forget use thermal paste). Did you check you have the right jet plate installed too, in the manual there are different plates for different cpu's so make sure you have the plate for 2066 cpu installed rather than a 1151 by mistake.

with thermal compound too try and get a nice even spread across the whole ihs, try the cross technique but make sure the whole ihs is covered

 

[Stilez]

regarding the instructions for block installation the orientation of the block can have a massive impact on cpu temps, you need the base plate of the block where the fins are cut into the metal ideally parallel to the cpu die direction that way you maximize the cooling area of the block relevant to the cpu, to find that out you need to take the block apart and see which direction the fins are pointing.

if the base plate fins are in a vertical position that is the wrong way, as a small area will make contact with the ihs and directly under the die, you need the base plate fins to be in a horizontal position as that way the whole ihs is being cooled down by the block, you'd see better load temps.

if the above has you confused don't worry, the long and short of it watercool's manual isn't very clear on which way the block should install, i still recon that it's not installed correctly and rotating the block 90 degrees to the right will likely help (take the block off and turn it 90 degrees to the right and re install it, don't forget use thermal paste). Did you check you have the right jet plate installed too, in the manual there are different plates for different cpu's so make sure you have the plate for 2066 cpu installed rather than a 1151 by mistake.

with thermal compound too try and get a nice even spread across the whole ihs, try the cross technique but make sure the whole ihs is covered

Photos in review articles clearly show which way the fins run in my block. Right now they're indeed vertical. I would have expected the CPU die was central enough or square enough for orientation not to matter but if it's large + rectangular, then I can see the point you're making. I'll try horizontal not vertical for sure. One thing, are you sure that horizontal would be the usual/expected Intel HCC die orientation? I couldn't find a schematic photo of this CPU, or how.large the die is compared to the IHS.

The block doesn't have jet plates (the manufacturer says), or rather they're part of the block top and not detachable.

 

[wookiee87]

Photos in review articles clearly show which way the fins run in my block. I would have expected the CPU die was central enough or square enough for orientation not to matter but if it's large + rectangular, then I can see the point you're making. I'll try horizontal not vertical for sure. One thing, are you sure that horizontal is the usual Intel HCC die orientation? I couldn't find a schematic photo of this CPU, or how.large the die is compared to the IHS.

The block doesn't have jet plates (the manufacturer says), or rather they're inbuilt in the block top. One thing less to go wrong

Intel cpus usually have a rectangle die that sits in the vertical position in the socket, would be similar on 2066 but given the size of the die it would more square but still rectangleish in size so waterblock install would be critical to get it right to achieve good temps, regarding the block it's very odd that they don't have jet plates, but i use ek stuff so other manufactures may have different ways to cool like watercool.

have a go at rotating the block and see what happens, i'd say your temps should drop but don't expect sub 60 degrees under load, after all its a 18 core at 4.4ghz+, so she'll be a hot one for sure

EDIT: what voltages are you running, you said the chip can run from 4.4 - 4.5ghz but at what voltage?

 

[Stilez]

EDIT: what voltages are you running, you said the chip can run from 4.4 - 4.5ghz but at what voltage?

With all cores at 4.5 GHz, vcore varies from 1.070 to 1.230. Thats a guess, not definitive, still stress testing and discovering voltages needed, but best guess until I'm done. For example, right now, one core will do 47x stable at 1.140, another core failed 45x at 1.210 (still testing so no.definitive final values)

 

[wookiee87]

With all cores at 4.5, vcore varies from 1.070 to 1.230. Thats a guess, not definitive, still stress testing and discovering voltages needed, but best guess until I'm done. Right now, one core will do 47x at 1.140, another core failed 45x at 1.210 (still testing so no.definitive final values)

with the 18 core chip you have its advisable not to break 1.25v and in the very best circumstances keep it around 1.165v, i'd head in and set your multiplier to 44/45 and a manual voltage of around 1.18v, if the pc is stable, drop the voltage to around 1.15v and see if it's ok, if so leave it at that, likely the auto voltages supplied are too much for the cpu, and the cores that are boosting higher are causing the cpu to heat up.

did you change the orientation of the cpu block?

 

[Stilez]

with the 18 core chip you have its advisable not to break 1.25v and in the very best circumstances keep it around 1.165v, i'd head in and set your multiplier to 44/45 and a manual voltage of around 1.18v, if the pc is stable, drop the voltage to around 1.15v and see if it's ok, if so leave it at that, likely the auto voltages supplied are too much for the cpu, and the cores that are boosting higher are causing the cpu to heat up.

did you change the orientation of the cpu block?

I haven't yet reoriented the CPU block, I'm middle of stress testing so I'll do that afterwards, tonight or tomorrow and compare temps.

I agree totally, the built in voltages are way too high. But also, with this chip (maybe for all HCC chips?) it looks like the variability between cores means that it's working better to set Vcore on a core-by-core basis, so the few cores needing more voltage or generating more heat don't drag down the rest.

A few quick tests suggest I can get 45x on all cores 24/7 stable at an acceptable temperature (for stress testing that is). I've started by dialling a fixed 45x on all cores and a starting guesstimate of 1.140v. Then prime95 non-AVX, and increase cores in 10mv steps as they show errors, until all cores seem to run error-free for long periods (12-24 hrs: I've rarely had some cores error out after only 4-6 hours). Before watercooling I was struggling to get 43x stable and pushing TJmax - 5 C. I'm currently getting 45x at about 89-90 C. That's not a great temp for 24/7 use but for stress testing its okay.

Right now, midway through testing stress, speeds, voltages and heat, this is the position at 45x all cores and 32x mesh:

<=1.050v … 3 cores (still testing 10mv at a time how low these cores can go)
1.060v … 3 cores
1.070v … 2 cores
1.100v … 2 cores
1.130v … 1 core
1.160v … 2 cores
1.170v … 1 core
1.180v … 2 cores
1.200v, 1.220v ... 1 core each

Once i get 45x 24/7 stable, those cores that are low on voltage and heat might be pushed to 46x ~ 48x. But I am feeling hopeful about getting 45x all cores stable on prime95 maxxing out at < 90 C worst case 24/7 stable. If I can, I'll be fairly happy, 45x is pretty typical for this CPU according to silicon lottery binning. If not, it'll be a few at 44x and the rest 45-48x, I shouldn't need to go lower.

While I haven't pushed other things hard, its looking like mesh 30x at 150mv uplift (and maybe 32x @ +250mv) may also be stable. At least, Im running that and so far no memory errors obvious.