CompuLab Turns An 8-Core/16-Thread Xeon, 64GB RAM, NVIDIA Quadro RTX 4000 Into Fan-Less Computer

[patrakov]

This offering isn't so special after all - you can get all kinds of passively cooled systems, including ones with Threadripper CPUs and high-end GPUs, also at https://www.deltatronic.de/en/pc-en

I own two of their tower systems, the first one was bought when the Athlon64 was new on the market, and I am pretty satisfied with how well they work.

The fact that there is an alternative is important, especially because some countries (e.g. Malaysia) ban all imports from Israel.

[pegasus]

Modern server vendors rely on fresh air coming up through the bottom of the rack and taken in en masse through the front panel where all system generated heat is pulled through the back and pushed up through the top.

...

Also for maximum effect the Compulab has to be installed vertically, where rack servers are optimized to be installed horizontally.

Yes, that's the thing. Modern rack is designed for comfort, not for energy efficiency since it needs to spend energy to push air where it does not want to go. Arranging servers in the rack vertically would allow for natural convection, air circulation being driven by heat generated by servers itself, getting rid of the fans and reducing power consumption by something like 20-30%.
I kinda hoped that OCP and similar projects would come up with such a design, but it seems that horizontal placement is valuable enough that it is cheaper to lay down another plumbing and invent dry water disconnect plugs rather than using natural effects to cool stuff down.

At least designs such as Sugon Nebula take advantage of convection, but in liquid

[coder]

Pretty impressive results for a decent specced system, especially considering it isn't that large.

How can you say that without seeing any benchmark results or at least a graph of the CPU & GPU clocks?

All I need to see are the power graphs, in order to know that this thing is configured to aggressively down-clock, to keep the thermals in line.

[coder]

I did NOT mean keeping them (the 'innards') locked up in a sealed box and merely blowing a fan ONTO the sealed box (which would do absolutely nothing to cool the sealed-up contents of the box).

This is incorrect. By increasing cooling of the outside, you're increasing the thermal gradient, which improves thermal conductivity, thereby reducing temperatures inside the box.

Wouldn't be at all surprised to find that an 'open system' without a fan would result in a considerable reduction in overall temperature(s).

This has been demonstrated, repeatedly. I've even informally observed it on a few occasions. Most recently, a coworker's PC was running a heavy multi-core workload, causing its CPU fan to spin up to the point where it bothered me. I cracked open the side-panel and it quieted down substantially.

[coder]

Modern rack is designed for comfort, not for energy efficiency since it needs to spend energy to push air where it does not want to go. Arranging servers in the rack vertically would allow for natural convection, air circulation being driven by heat generated by servers itself, getting rid of the fans and reducing power consumption by something like 20-30%.

And these numbers are from where? Convection will not drive nearly enough airflow, for modern servers. It might be marginally easier for the fans to work with convection, but I doubt to the degree of delivering such a large energy savings.

Also, how are you going to achieve the same level of density (in terms of CPU cores per m^2 of floor space), with vertically-oriented boards?

I kinda hoped that OCP and similar projects would come up with such a design, but it seems that horizontal placement is valuable enough that it is cheaper to lay down another plumbing and invent dry water disconnect plugs rather than using natural effects to cool stuff down.

At least designs such as Sugon Nebula take advantage of convection, but in liquid

Lately, full immersion cooling has been getting rather popular, in HPC.

[pegasus]

Numbers come from server vendors I'm talking to. Those fans can easily go over 30W per fan at full power and your typical pizzabox has 8-12 of them.
Blade servers are already aligned vertically, but still have airflow front to back. Some major rethinking would be needed for a bottom to top airflow design and careful arrangement of components inside such rack would be needed (most heat sensitive at the bottom, least heat sensitive at the top), but I'm sure that on a rack level energy savings would be substantial.
There are companies already exploring this direction, see for example vapor.io/chamber .

[ruthan]

I would like to something beefier eith faster gpu geforce 1660 and pvepriced quadro are slimmers. I backed Calyos fanless project bit it seems that i throw money out of yhe window..

[schmidtbag]

How can you say that without seeing any benchmark results or at least a graph of the CPU & GPU clocks?

All I need to see are the power graphs, in order to know that this thing is configured to aggressively down-clock, to keep the thermals in line.

Because we already know roughly how good that CPU and GPU performs. It was very clear that they weren't thermal throttling.

[AsuMagic]

Because we already know roughly how good that CPU and GPU performs. It was very clear that they weren't thermal throttling.

Really? I am pretty sure that at a 96°C maximum, the CPU did thermal throttle.
I count in not reaching the max turbo speed due to thermal constraints as thermal throttling personally, not just throttling below the base clock.

[schmidtbag]

Really? I am pretty sure that at a 96°C maximum, the CPU did thermal throttle.
I count in not reaching the max turbo speed due to thermal constraints as thermal throttling personally, not just throttling below the base clock.

Fair enough, though, it didn't peak that high for very long either. I don't think most people who get a PC like this are going to push the CPU under 100% load for extended periods of time.