Has anyone considered water cooling a spark board? I have been wondering what the performance and reliability differences would be with a CNC water-cooling jacket.
I have been in discussions with some CNC service providers to design something.
Any opinions on this?
I think it will reduce the thermal throttling but it voids the warranty and there is always risk of water leak.
A complete solution for DGX spark might be including a redesign housing to reduce dust and Foreign Object affecting the PCB.
Redesigned husing offer some extra, it might make 2280 ssd practical.
A Mad Max/steampunk design might be attractive. I will throw this into ChatGPT Image 2 and see if anything attractive come out.
here is a solution for the 5090 as an example:
This looks like a temporary solution - and, as we know, temporary solutions often become permanent. =)
However, the ASUS radiators will quickly get clogged with dust, and the airflow will become much weaker than the manufacturer originally intended. I recommend thinking about how to filter the dust so it doesn’t get pulled into the ASUS itself. Right now, it seems like the ASUS is already sucking in dust, and the additional fans are pushing even more dust into it.
Alternatively, it will need to be disassembled and cleaned from time to time.
Agreed this was just to see if keeping the unit cooler actually makes a difference. And it does, the sudden death power downs have stopped.
Now I need to figure out a permanent solution to run a cluster of 8 because this is not optimal!
I am building custom water cooling loops for many years now. It’s not that hard and I had never a leaking loop. There ist not much pressure in the system. Therefore minimal risk, if you do it right and use proper connectors. There are many brands available.
I am also tempted to water cool my sparks …
There is a main issue with all water coolers. First you have to define which parts have to be cooled. I am pretty sure the Nvidia cooler is also cooling the LPDDR5 chips and probably the voltage regulators of the GB10- Usually they come in different heights, therefore the watercooling block will need a 3D shape to cover all the parts. Next you have to choose, if you want to use thermal putty, thermal pads or heat transfer paste to bridge all distances between the parts and the cooler. That’s the tricky part. If e.g. the pads on the voltage regulators are slightly to thick, the GB10 might have a gap between the cooler and the silicon die. That might be the end of the GB10 (not really, because usually these parts are throttling before they overheat, but it’s still a risk)
I would consider building a waterblock that really fits far to much effort. There is only one person I know who might be interested to design such a cooler. Roman (der8auer) Hartung with his company “thermal grizzly”. If you are really interested he might be the one to ask. He could also manufacture the coolers. If we could find some people who are willing to pay for such a cooler, we could split the development costs …
Next solution would be to piggy-back a water cooler on the existing air-cooler. I don’t know if that’s possible without machining the original cooler. Does anyone have a picture of that cooler? To improve the cooling, one could use a Peltier element between the two coolers, because cooling a cooler is not very efficient. The downside is: the peltier element itself consumes up to 50W which also has to be dissipated by the water cooling cycle. Do a research of that principle. Some companies tried that solution for CPU coolers in the last decades. It’s working but for a standard CPU cooler it’s overkill. The Peltier elements are affordable standard electronic components.
Don’t know if that helps …
So for me options ar:
Optionally:
I need this urgently and will meet with the manufacturer of the system shown in the attached images.
I just asked for a quote at Thermal Grizzly. Development cost plus cost per piece for a volume of 40 units.
Very quick reply … they are not interested
No capacity, too much effort, too costly.
@Brush China it is then.
Let’s get some replies on the units of interest. I will be taking 8 to 16 units. Any other interested parties?
Put it on kickstarter might get you more people to join. I will ask around in Taiwan but no promise and it is not going to be fast. We probably have people from China in this forum.
Mate, no promises in this world… but maybe a joint kick starter?
And to folks in China, if any of you are in ‘Luancheng District, Shijiazhuang City, Hebei Province, China’, maybe we can meet up for a meal and beers and chat Aayeei ?
I will make some phone call today and we will see what I got.
I have been working in Taiwan computer supply chain for 30 years so there is still some connection I can check. One of them is willing to discuss. This project will require engineer who can do mechanical drawing and talk to CNC workshop. I have the ASUS GX10 so I will exam it with that company tomorrow.
My plan is getting an estimation of the cost and possible specification. Anyone who is interested are welcome to join discussion.
If there is enough interest, maybe we can setup a discord group chat.
My plan is creating a waterblock solution and enable DGX Spark to accommodate 2282 SSD.
Turn them upside down, keep the rest the same. You should see another slight drop in temps :)
Turning upside down resulted in more frequent power-offs, and a faster time to power down
@paulsc.liu I have a unit with a CNC company already.
NvidiaGB10WaterBlock for Discord group
