Most active commenters
  • Dylan16807(3)

←back to thread

Google's Liquid Cooling

(chipsandcheese.com)
399 points giuliomagnifico | 15 comments | 25 Aug 25 17:57 UTC | HN request time: 1.372s | source | bottom
1. michaelt ◴[25 Aug 25 18:51 UTC] No.45017512[source]▶
> TPU chips are hooked up in series in the loop, which naturally means some chips will get hotter liquid that has already passed other chips in the loop. Cooling capacity is budgeted based on the requirements of the last chip in each loop.

Of course, it's worth noting that if you've got four chips, each putting out 250W of power, and a pump pushing 1 litres of water per minute through them, water at the outlet must be 14°C hotter than water at the inlet, because of the specific heat capacity of water. That's true whether the water flows through the chips in series, or in parallel.

replies(4): >>45017544 #>>45017835 #>>45017973 #>>45021855 #
2. foota ◴[25 Aug 25 18:54 UTC] No.45017544[source]▶
Hm... but in the case when the chips are in serial, the heat transfer from the last chip will be less than when the chips are in parallel, because the rate of heating is proportional to the difference in temperature, and the water starts at a lower temperature for the parallel case for this last theoretical chip.
replies(3): >>45017827 #>>45018696 #>>45018736 #
3. fraserphysics ◴[25 Aug 25 19:20 UTC] No.45017827[source]▶
One way to characterize the cost of cooling is entropy production. As you say, cooling is proportional to difference in temperature. However, entropy production is also proportional to temperature difference. It's not my field, but it looks like an interesting challenge to optimize competing objectives.
4. idiotsecant ◴[25 Aug 25 19:21 UTC] No.45017835[source]▶
It just means in series that some of your chips get overcooled in order to achieve the required cooling on the hottest chip. You need to run more water for the same effect.
replies(1): >>45017998 #
5. friendzis ◴[25 Aug 25 19:33 UTC] No.45017973[source]▶
While there is some truth to your comment, it has no practical engineering relevance. Since energy transfer rate is proportional to temp difference, therefore you compute the flow rate required, which is going to be different if the chips are in series or in parallel.
replies(1): >>45034315 #
6. k7sune ◴[25 Aug 25 19:34 UTC] No.45017998[source]▶
I can imagine a setup where multiple series of slower cooler water converging into a faster warmer stream, and the water will extract an equal amount of heat away from all the chips whether upstream or downstream.
7. 0x457 ◴[25 Aug 25 20:39 UTC] No.45018696[source]▶
Yes, but water is constantly moving in a loop. It's not like you use water to cool chip #1, and then it moves to chip #2, it's constantly moving, so temperature delta isn't that much.
replies(1): >>45019066 #
8. chickenbig ◴[25 Aug 25 20:43 UTC] No.45018736[source]▶
In steady state the power put into the chip is removed by the water (neglecting heat transfer away from the cooling system). The increased water temperature on entering into the cooling block is offset by a correspondingly higher chip temperature.
9. smachiz ◴[25 Aug 25 21:10 UTC] No.45019066{3}[source]▶
in their first serial design, that's exactly what it was doing.
replies(1): >>45034330 #
10. JonChesterfield ◴[26 Aug 25 03:08 UTC] No.45021855[source]▶
The equations all work out OK though since one can go substantially above a litre a minute at pressures which are still viable. The standard 18W thing for desktops is about ten times that.
11. Dylan16807 ◴[27 Aug 25 01:12 UTC] No.45034315[source]▶
If you're heating up the water by 10 or so degrees on typical computer hardware, I bet you're not bottlenecked by energy transfer. Your flow rate is based on how hot you want the water to get, so series or parallel goes back to not mattering.
12. Dylan16807 ◴[27 Aug 25 01:14 UTC] No.45034330{4}[source]▶
They mean it doesn't sit around and wait until it's hot, which then makes it "used" and unable to cool something else. It keeps moving through the first chip and is only a small percent saturated.
replies(1): >>45044451 #
13. smachiz ◴[27 Aug 25 20:04 UTC] No.45044451{5}[source]▶
depends how fasts it is moving - but it's all moving at the same speed. The last chip will get less cooling than the first chip, proportional to the amount of cooling the previous chips received. Delta T doesn't lie.

Whether it's significant or not, I can't know - but you want it to be significant, otherwise it's less efficient.

replies(2): >>45044755 #>>45120066 #
14. Dylan16807 ◴[27 Aug 25 20:27 UTC] No.45044755{6}[source]▶
> depends how fasts it is moving - but it's all moving at the same speed. The last chip will get less cooling than the first chip, proportional to the amount of cooling the previous chips received. Delta T doesn't lie.

The watts of cooling per chip will be the same. The last chip in the loop will be a little warmer, but not by much in a reasonable setup. The difference in temperature between each chip and the water running across it will be the same.

And if you take a weak water supply and then split it to run in parallel, you can end up with a significant heat gradient across each waterblock which doesn't sound great either. If you have 4 high power chips please don't limit them to .25 liters per minute each.

> Whether it's significant or not, I can't know - but you want it to be significant, otherwise it's less efficient.

Keeping your fluid cool is good for long term reliability. And if you're doing that, then every block is getting cool fluid and the other details about loop layout won't matter.

15. 0x457 ◴[03 Sep 25 20:30 UTC] No.45120066{6}[source]▶
It moves fast enough that the temperature delta between first and last chip isn't significant. Not too fast, though, because if the pump is running full-tilt it becomes a major heat-source itself.

> but you want it to be significant, otherwise it's less efficient.

No? You want a large delta between heat-source and water, as well as water and whatever is on your cool side. Essentially just between hot and cold sides.

Delta between heat-sources is irrelevant unless you start cherry-picking some very odd combinations.