Water cooling of 0.5-1 kW processors

[nexox]

I don't know exactly what's available but I have noticed that window units and mini splits have gotten much more efficient and cheaper in the last few years, I tend to assume that carries over into portable units, but maybe not.

I searched briefly and found some random 8000BTU Frigidaire unit that claims a combined efficiency of 6.3, which means that the ~800W of power consumption would pull ~4800W of heat out of the room, so you wouldn't have to run it constantly. Getting something like that to switch on and off when you want it may not be so easy.

 

[DanRR]

I don't know exactly what's available but I have noticed that window units and mini splits have gotten much more efficient and cheaper in the last few years, I tend to assume that carries over into portable units, but maybe not.

I searched briefly and found some random 8000BTU Frigidaire unit that claims a combined efficiency of 6.3, which means that the ~800W of power consumption would pull ~4800W of heat out of the room, so you wouldn't have to run it constantly. Getting something like that to switch on and off when you want it may not be so easy.

Do you mean SEER/EER 6.3 or COP ? If SEER then the COP efficiency is less than 2.

 

[nexox]

SEER is some other number that I haven't bothered to understand, and it's usually somewhat higher, combined efficiency is the heat pump COP (this is just a ratio of power in to heat moved, easy to understand) minus some other power consumption, probably the fan and electronics.

 

[DanRR]

I misunderstood, was it COP 6.3? Can you post the link on that miracle 800W air conditioners which will transfer 4800W of heat, may be I missed some great invention lately

 

[nexox]

Can post the link on that miracle 800W air conditioners which will transfer 4800W of heat

That's not a miracle, that's a relatively low efficiency because the unit is fairly small, it's just that 4500W of heat transfer isn't really all that much as far as changing the temperature of a room. The specs could of course be a little off because in reality those things vary with conditions, but like I said, these things have been getting more efficient.

 

[RolloZ170]

Can you post the link on that miracle 800W air conditioners which will transfer 4800W

my CPU eats 350W and the Cooler FAN needs 12Watts to cool 350W away.

 

[DanRR]

my CPU eats 350W and the Cooler FAN needs 12Watts to cool 350W away.

Yea, kilowatts is where we all heading too. The Arctic fans cooled the radiator almost to the same room temperature and the server processor is still at dangerous temperatures. Basically we reached the limit. So the claim at the beginning of this thread that super-pooper gazillion RPM fans will solve the problem for the next several generation of processors is dead. Further reduction needs changing room temperature which is also too high due to high consumption and at summers it will be beyond the hell

 

[nexox]

I would look at increasing the coolant flow rate through the water blocks first, though I don't know about how those specific water coolers are set up, how that could be achieved, or how much pressure those systems can handle.

 

[RolloZ170]

The Arctic fans cooled the radiator almost to the same room temperature and the server processor is still at dangerous temperatures

bad thermal paste, too much paste. waterflow too few.
hotspot of processor at wrong place / wrong waterblock type.

 

[RolloZ170]

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.

this is not correct.
a Xeon W9-3495X overclocked and consuming 900Watts was cooled with Noctua NH-U14S DX-4677
( two FANs at room temp ) very extreme but 600Watts should be doable.
CBR23 500 watts, about 63 degree temps, compared with Custom loop 360 radiator 52 degree temp.

if you can't manage this with water something is realy wrong.
edit: i saw a video from computex where enermax showed a system with OCed W9-3495X consuming 900W cooled with Enermax LiqTech XTR 360

.

 

[CyklonDX]

edit: i saw a video from computex where enermax showed a system with OCed W9-3495X consuming 900W cooled with Enermax LiqTech XTR 360

I could see it being able to manage running a benchmark for a minute.
I don't see it being capable taking on the load for longer like a compute/render job - that can take even few days of 24x7 work. I presume most people are ok when your cores downclock to manage the temps.

 

[RolloZ170]

I could see it being able to manage running a benchmark for a minute.

it runs the whole day there.
but if only a minute, a std. 360 AIO with weak FANs cools 900W and i can't imagen there is a problem with 500-600Watts.

 

[CyklonDX]

yeah it runs a benchmark once in a while, cools down *(didn't even warmed up the fluid in the loop), and idles 90% of time there; when it idles, and its likely around 100W or even lower. I don't see a problem either if its just some games, or idling server with occasional load for a minute or 2.
(they have capacity to soak for couple minutes, but their actual cooling performance is much lower on the 240mm aio)

// I have a problem to cool down Xeon's 6527P's on full load without dropping clocks significantly in a R770 poweredge servers with fans at full speed going to 25k rpm; and thats 255W part with heatsinks rated for 300W - in datacenter with 10'C intake. *(being written 255W on intel site, in real life it feels like 500-600W at full throttle)

 

[RolloZ170]

my guess is the waterblock has wrong design for CCD positions of SP5 EPYC,
it seems the water gets not even warm.

 

[CyklonDX]

my guess is the waterblock has wrong design for CCD positions of SP5 EPYC,
it seems the water gets not even warm.

idk about that - but with any watercooling it takes a while to warm up water/fluid. Its what i meant water cooling allow you to soak heat (a buffer); but you are only limited to what your heatsink/radiator can cool. The 5minutes you run your sets of benchmarks will likely only heat your fluid by 1-5'C if its moving, if you run any waterloop for an hour you will hit a wall at some point because you can't dissipate it fast enough or/and room where you have radiator will warm up too.)

*if cpu says 500-1000W that's insane to cool down at full load on those parts - we are reaching levels of insanity - and those ai datacenters that will use it to 100% use crazy cooling solutions for them.

 

[nexox]

If the radiator is running at room temperature and the CPU is overheating then there's clearly a different issue than short term heat capacity here, either insufficient flow rate or the water block is not transferring the heat out of the CPU adequately.

 

[DanRR]

That was still too early, preliminary data. Also preliminary, the difference between the room temperature and the sensors with highest reading are 41C plus some fluctuations give 1-3 degrees Celsius more.

The room temperature is pretty high around 33C or more despite some not efficient and not optimized cooling from the outside. At night it is 28-29C.

More or less final is that changing own Silverstone XE360 fans to Arctic 4k gave around 5 C drop in processor temperature. Around 6 C drop was obtained changing just one stock fan to Arctic 8k. Next will be taking two Arctic 4K+one 8K when computer will be less busy.

 

[RolloZ170]

but you are only limited to what your heatsink/radiator can cool.

the best one will not work if the waterblock can't get the heat.
wrong pressure.
too much/few/wrong thermal paste
hot spot wrong place. remember at SP5 Turin the heat comes from outer sides not central.

 

[RolloZ170]

More or less final is that changing own Silverstone XE360 fans to Arctic 4k gave around 5 C drop in processor temperature. Around 6 C drop was obtained changing just one stock fan to Arctic 8k. Next will be taking two Arctic 4K+one 8K when computer will be less busy.

sorry, we may have overseen something.
Silverstone can cool LGA4677 Xeons easy because they have more central hot spot.
AMD EPYC SP5 have springs for CPU pressure, they may not give optimal pressure but you can't change anything here.
can you explain/show your procedure of insatlling the waterblock pls.
which thermal paste, how much (more is not better) and your applying pattern.
maybe less paste of the best thermal paste will give better results.
i think an waterblock optimized for Turin CCD placement is required because you can't change pressure of.
sorry.

 

[DanRR]

sorry, we may have overseen something.
Silverstone can cool LGA4677 Xeons easy because they have more central hot spot.
AMD EPYC SP5 have springs for CPU pressure, they may not give optimal pressure but you can't change anything here.
can you explain/show your procedure of insatlling the waterblock pls.
which thermal paste, how much (more is not better) and your applying pattern.
maybe less paste of the best thermal paste will give better results.
i think an waterblock optimized for Turin CCD placement is required because you can't change pressure of.
sorry.

All coolers designed for SP5 socket of course were not designed for Turins. I used two: the air cooler resembling but not from this company


and Silverstone XE360-SP5. They are made for Genoas. They all came with thermopaste on the surfaces. On Genoas this air CPU cooler had better results than other people reported with other air coolers. Still the Silverstone waterblock gave ~10C smaller temperatures than on air CPU cooler, and when moved to Turin the temperature increased by ~10 C as power consumption jumped by 450-550W. The temperature is monitored all the time by Psensor app (sensor called Tctl). Since that I changed waterblock many times between Genoas and Turins. The MX-4 or MX-6 thermopaste gave no noticeable differences versus original pastes or each other.

Due to all that I find the whole this picture very reasonable.

To overturn this picture the actual results from the Turin or high end Genoa with 96 or 128 cores are needed running all cores, ideally with floating point not just integer and the Silverstone cooler