Water cooling of 0.5-1 kW processors

Notice: Page may contain affiliate links for which we may earn a small commission through services like Amazon Affiliates or Skimlinks.

DanRR

Member
Feb 4, 2024
84
6
8
Any ideas how to cool SP5 socket processors like EPYC Turin which dissipate 0.5-0.6 kW ? Air coolers already do not work there. Water coolers like Silverstone360 designed for Genoa EPYC processors with TDP below 400W can not keep up with this type of power with temperatures to reach 80 C. And coming EPYC Venice processors with future SP7 socket expected to drain even larger power. Looks like this crazy power race will never stop.

As a quick solution - is it possible to upgrade current water coolers to work in parallel? Computer cases like CORSAIR 9000D have double or even triple and quadruple space for coolers
 
Last edited:

siematos

Member
Dec 7, 2021
46
32
18
Any ideas how to cool SP5 socket processors like EPYC Turin which dissipate 0.5-0.6 kW ? Air coolers already do not work there. Water coolers like Silverstone360 SP5 designed for Genoa EPYC processors with TDP below 400W can not keep up with this type of power with temperaturess to reach 80 C. And coming EPYC Venice processors with future SP7 socket expected to drain even larger power. Looks like this crazy power race will never stop.

As a quick solution - is it possible to upgrade current water coolers to work in parallel? Computer boxes like CORSAIR 9000D have double or even triple and quadruple space for coolers
first of all, it's a watts-per-area type of issue, where you have a small area dissipating a ton of heat. The limiting factor here is the material it needs to transfer through, like copper or thermal paste. If you reduce those layers or improve the thermal interface, you can get a lot more oomph out of your cooling solution (here a link). Another approach is through chip design/manufacturing: If you remove the amount of silicon (which has a bad heat transfer characteristic) between the parts that produce the heat and the heatsink, you will improve heat dissipation. This should result in a smaller difference between the hottest parts/junctions and the coldest parts, which allows for the entire chip to run a bit hotter overall. The next step then would be direct cooling, either direct or indirect. There was an article about this on STH a few weeks ago. While this is a water cooling solution, you could make something work with a vapor chamber design as well in place of the heatspreader we got currently. That's more of a packaging technology challenge, but could produce socket'able CPUs close to the design we have right now. all of the latter solutions would get us comfortably into the 1kw+ territory in terms of potential cooling, and it's more an issue of reducing the costs. You can get really good and efficient designs as well. It just stands in a direct conflict with peak performance due to physics/manufacturing capabilities, since you can only push design so far. It's not a crazy power race. It's a crazy performance race, and efficiency for compute is rapidly rising because of it. We don't want it to stop, it's a good thing! And in terms of cooling, we're at a fraction of what's possible. We're just balancing it against what's economical.
 
  • Like
Reactions: DanRR

i386

Well-Known Member
Mar 18, 2016
4,801
1,863
113
36
Germany
Put 10k+ rpm 120mm fans on that single 360cm radiator and it will easily cool the next 3 generatiosn of the largest epyc cpus :D
(but these fans will bring their own problems: noise, power, weight)
 
  • Like
Reactions: siematos and DanRR

TrashMaster

Member
Sep 8, 2024
37
30
18
  • Like
Reactions: DanRR

DanRR

Member
Feb 4, 2024
84
6
8
Thanks for suggestions.
Looks encouraging. How much decrease of temperature you got? And have you tried Noctua NF-F12 iPPC 3000 PWM, Heavy Duty cooling fan, 3000 RPM? I have one, it works like a power horse. It has 15 years MTTF, this Arctic 6 years
 
Last edited:

TrashMaster

Member
Sep 8, 2024
37
30
18
Thanks for suggestions.
Looks encouraging. How much decrease of temperature you got? And have you tried Noctua NF-F12 iPPC 3000 PWM, Heavy Duty Cooling Fan, 3000 RPM? I have one, it works like a power horse. It has 15 years MTTF, this Arctic 6 years
I have tons of Noctua stuff sitting unused on my "only fans" shelf. Due to price/performance (when my larger projects might require 30+ fans) I am now an Arctic fan-boy. (ha ha ha i am very punny on a caturday morning.)

This is where with fans, thickness improves the static pressure:

Noctua: (25 mm thick)
Airflow 109.89 CFM
Static Pressure 7.63 mm H₂O

Arctic: (38 mm thick)
Airflow 106 cfm
Static Pressure 11.45 mmH2O

Their airflow might be comparable but the static pressure on the Arctic is what you need to really drive a fine finned rad hard.

LATE EDIT:

Because there were questions about the stock aio fans from silverstone here are their specs for comparison as well:


Speed: 600 ~ 2800 RPM
Rated Current: 0.25A
Max airflow: 87.72 CFM
Max air pressure: 3.09 mmH2O
 
Last edited:
  • Like
Reactions: DanRR and jode

TrashMaster

Member
Sep 8, 2024
37
30
18
That is a complex question to answer given the chips I am using in this configuration.

The thermal variance across a large chip like the lga 4677's in my video varies wildly, and under load will consume all available power then thermal throttle.

So my tests were how much performance could you squeeze out of the chip before all cores clocked down at 90c.

Best results were by far the Arctic 4k 120mm server fans. Going to the faster Arctic fans yielded only nominal improvement because the radiator was not thick enough to result in additional thermal exchange, and then going totally custom the cold plate itself becomes a bottleneck.
 

Attachments

  • Like
Reactions: DanRR

DanRR

Member
Feb 4, 2024
84
6
8
That is a complex question to answer given the chips I am using in this configuration.

The thermal variance across a large chip like the lga 4677's in my video varies wildly, and under load will consume all available power then thermal throttle.

So my tests were how much performance could you squeeze out of the chip before all cores clocked down at 90c.

Best results were by far the Arctic 4k 120mm server fans. Going to the faster Arctic fans yielded only nominal improvement because the radiator was not thick enough to result in additional thermal exchange, and then going totally custom the cold plate itself becomes a bottleneck.
Can you clarify a bit more this plot? Does it mean that baseline fans were at 90 C and Noctua dropped it to middle 60th C or 30C down? Or Stock was at 70 C and Noctua dropped it to 64.2C ?
 

DanRR

Member
Feb 4, 2024
84
6
8
And have you tried same size, just one dollar more expensive, monstrous power, 5x the 4K
ARCTIC S12038-8K ?
 
Last edited:

DanRR

Member
Feb 4, 2024
84
6
8
If using high power fans like ARCTIC S12038-8K consuming 25W each, 150W total for all 6 fans, it is necessary to power them from the external 12W source because motherboard most probably will burn while doing that. The motherboard CPU Fan ports in this case should be used as a voltage regulators of external power supply. Any reference on that kind of design? For sure it was done before.
 

DanRR

Member
Feb 4, 2024
84
6
8
Please keep STH G rated. :)

But air cooling and 500W is just a fan airflow issue on Turin, not an issue like it cannot be done.
500W Turins already one year in the market but surprisingly so far there is nothing in the market to cool them. It's time to encourage producers to make powerful coolers. They will benefit previous generations processors too, because they will last longer, throttle less and make overclockers happier
 

DanRR

Member
Feb 4, 2024
84
6
8
25W max. not all the time. you want CPU cool or not ?
motherboards are made for that. +12V comes from PSU, PWM says the FAN what todo.
Cooling needed not for spikes of power, but for many hours/days constant running
 

DanRR

Member
Feb 4, 2024
84
6
8
4pin have +12V all the time, not regulated, don't mix with 3pin FANs.
Super, I did not know that, still isn't it too much for 4pin port to drain 6A current /75W ? Or it is possible to offload power to PSU 12V ports (making adapter to new 4pin port) and take one controlling wire from the motherboard 4pin to new 4pin?
 
Last edited: