passive cooling [Finished] The number cruncher: Triple Xeon passive mineral oil cooling [Update 12: Final pictures and summary]

[Trey222]

It is amazing what this older hardware can do. Especially for the price. 

[Stefan1024]

17 hours ago, thekeemo said:

ETA for oil?

Several month. The base plate is not even ordered jet, so all the oil would spill out.

Also I have to wait for AMD / Nvidia to release the new GPUs.

 

17 hours ago, Trey222 said:

It is amazing what this older hardware can do. Especially for the price. 

Yes the compute power is really nice. But the power consumption to do so is massive. The Xeon Phi uses about 105 watts when beeing idle and doing nothing.

 

 

[Prysin]

you need to swap those Noctuas for IPPC fans.

Normal noctuas will not enjoy the resistance over time.

[Prysin]

5 minutes ago, Stefan1024 said:

The Xeon Phi uses about 105 watts when beeing idle and doing nothing.

seems like Intel could learn from AMD and Nvidia about "Idle power usage".....

[Stefan1024]

3 minutes ago, Prysin said:

you need to swap those Noctuas for IPPC fans.

Normal noctuas will not enjoy the resistance over time.

I will use oil with a low viscosity. They should be able to work in this condition. Or at least I hope so

When they fail, swapping them is not funny. But I don't like to buy fans for 60$ if I don't really have to.

1 minute ago, Prysin said:

seems like Intel could learn from AMD and Nvidia about "Idle power usage".....

They are not desiged to be idle. They are designed to be used in a cluster computer and running 24/7 under full load. When the data center don't need that much compute power, it shuts down some nodes completely.

[Prysin]

2 minutes ago, Stefan1024 said:

I will use oil with a low viscosity. They should be able to work in this condition. Or at least I hope so

When they fail, swapping them is not funny. But I don't like to buy fans for 60$ if I don't really have to.

They are not desiged to be idle. They are designed to be used in a cluster computer and running 24/7 under full load. When the data center don't need that much compute power, it shuts down some nodes completely.

but they're black..... they would fit into your color scheme.

[Stefan1024]

2 minutes ago, Prysin said:

but they're black..... they would fit into your color scheme.

I don't mine the brown color. Also you can't see the internals very well once finished.

[Prysin]

1 minute ago, Stefan1024 said:

I don't mine the brown color. Also you can't see the internals very well once finished.

we know its there Stefan. WE KNOW!!!

[Stefan1024]

Just now, Prysin said:

we know its there Stefan. WE KNOW!!!

Haha

Yes the IPPC once would look and perform better.

But..... I even had to overcome myself to use fans at all. The once I use now where bundled with the heat sinks so it 's ok. But buying fans separate? Meh.

Buying super noisy 3000 RPM fans? Noooooooo.

[Prysin]

8 minutes ago, Stefan1024 said:

Haha

Yes the IPPC once would look and perform better.

But..... I even had to overcome myself to use fans at all. The once I use now where bundled with the heat sinks so it 's ok. But buying fans separate? Meh.

Buying super noisy 3000 RPM fans? Noooooooo.

as if you gonna hear much of that noise through all that oil...

[Stefan1024]

Just now, Prysin said:

as if you gonna hear much of that noise through all that oil...

I have to be careful not to produce suface waves as they will emmit sound.

Also it's about breaking a principle.

[Prysin]

5 minutes ago, Stefan1024 said:

I have to be careful not to produce suface waves as they will emmit sound.

Also it's about breaking a principle.

the way you solve that is by using a fine metal mesh (dustfiler/medium fine mesh) across the surface. It will allow flow, but not waves.

[Stefan1024]

1 minute ago, Prysin said:

the way you solve that is by using a fine metal mesh (dustfiler/medium fine mesh) across the surface. It will allow flow, but not waves.

Good idea

[Prysin]

2 minutes ago, Stefan1024 said:

Good idea

actually, make that the "corse" hexa pattern mesh you find in cases. Reason is that fine dust filters will have too much drag when facing the viscosity of the oil. This will cause movement in the material, which can cause it to tear at the mounting points. which could cause metal mesh to float around your case. A sturdier, thicker metal mesh would have less flex, but also low enough drag coefficient that the oil shouldnt make ripples or waves, but rather just buldge out of the holes and spread out over the mesh before sinking back down.

[Stefan1024]

1 minute ago, Prysin said:

actually, make that the "corse" hexa pattern mesh you find in cases. Reason is that fine dust filters will have too much drag when facing the viscosity of the oil. This will cause movement in the material, which can cause it to tear at the mounting points. which could cause metal mesh to float around your case. A sturdier, thicker metal mesh would have less flex, but also low enough drag coefficient that the oil shouldnt make ripples or waves, but rather just buldge out of the holes and spread out over the mesh before sinking back down.

Yes this makes sense. I will makes tests to see if I need it at all. Also when I install the mesh you can hardly see througth and the colo sheme of the internals gets irrelevent

 

[Prysin]

Just now, Stefan1024 said:

Yes this makes sense. I will makes tests to see if I need it at all. Also when I install the mesh you can hardly see througth and the colo sheme of the internals gets irrelevent

 

We know stefan. WE KNOW.

 

#neverforget

[Trey222]

4 hours ago, Stefan1024 said:

Yes the compute power is really nice. But the power consumption to do so is massive. The Xeon Phi uses about 105 watts when beeing idle and doing nothing.

Yeah. That is one of the downsides of using server parts.  Have you thought about getting a redundant PSU or a USP. 

[nycesquire]

I've learned a lot from you, Stefan.  Looking forward to updates.

[Stefan1024]

19 hours ago, Trey222 said:

Yeah. That is one of the downsides of using server parts.  Have you thought about getting a redundant PSU or a USP. 

Despite I'm using a lot of old server gear, I don't actually build a server. So I don't need the extra reliability stuff and con save some money.

 

10 hours ago, nycesquire said:

I've learned a lot from you, Stefan.  Looking forward to updates.

Thank you, you are wellcome.

[Orangeator]

What exactly does the Co-processor do? Like how do you assign tasks to it? Can you run VM's on it? I honestly have never heard of an Intel Phi before. Can you spoon feed me how exactly one could utilize it?

[Stefan1024]

9 hours ago, Orangeator said:

What exactly does the Co-processor do? Like how do you assign tasks to it? Can you run VM's on it? I honestly have never heard of an Intel Phi before. Can you spoon feed me how exactly one could utilize it?

Think of it like a GPU with no video output. When a program is running on the CPU and has something large to compute, it can send the task to the GPU with openGL and the GPU does the heavy lifting and sends the results back.

But to do so the program has to support it. As Xeon Phis are not common in a PC, hardly any programm supports it natively. Also due to it's very different achitecture to GPU, it's can't be utilized by openGL.

However, some programms do use the intel math kernel library (MKL). When it get's a compute job by a programm and it set up correctly, it will automaticly offload it to the Phi.
As a result the programm doesn't need to support the Phi to make use of it. On the other hand the utilisation it quite low, as only smal parts of the work can be offloaded.

[pagani123]

9 hours ago, Stefan1024 said:

Think of it like a GPU with no video output. When a program is running on the CPU and has something large to compute, it can send the task to the GPU with openGL and the GPU does the heavy lifting and sends the results back.

But to do so the program has to support it. As Xeon Phis are not common in a PC, hardly any programm supports it natively. Also due to it's very different achitecture to GPU, it's can't be utilized by openGL.

However, some programms do use the intel math kernel library (MKL). When it get's a compute job by a programm and it set up correctly, it will automaticly offload it to the Phi.
As a result the programm doesn't need to support the Phi to make use of it. On the other hand the utilisation it quite low, as only smal parts of the work can be offloaded.

thats interesting i would think for the price (new) someone would want to go with a quadro or even tesla gpu. 

[Stefan1024]

53 minutes ago, Maybach123 said:

thats interesting i would think for the price (new) someone would want to go with a quadro or even tesla gpu. 

A GPU is a vector compute unit that is optimized for throughput. If your problem can not be (effectively) described with vectors or it's results are highly dependent on each other, you will have a hard time to get a reasonable speed up by adding a GPU. But with a Xeon Phi you can.

Also Xeon Phis are mostly found in data centers / super computer and there you optimize your application for the hardware it will run at. And there you get a high utilisation and perfromance. But if you have a vector based problem, using a GPU is the best you can do.

[thekeemo]

What software do you use that utilizes the phi?

[nmil]

Yet another awesome silent computing build. I don't quite get how the coprocessor works. So does it just act like another CPU or something? Could you maybe use this for ray tracing in CAD, cause that takes age for me.