Solving the cooling

How to cool down 4-5 KW of heat from the device packed into relatively tight place. My solution is air-cooling and precice airflow control. 


Forced Air Cooling

Forced Air Cooling

All 4 nodes are housed inside the 24U closed rack. Cold air is sucked in near the floor at the bottom and pushed to the back side of closed rack. Then its sucked from back to front through GTX 295 GPU cards. Hot air isimmeadetly  moved out from the rack, rising upwards to AC devices of the server room.

In the fan plate I use 3 x 170 mm diameter  240V AC fans at full speed.



  1. Great design.

    I just would like to pass my previous experience with large system cooling. Beacuse of pressure difference, naturally, heat builds an rises from bottom to upwards. Every large system’s bottom is cooler than its top. In your configuration, bottom nodes will have less heat stress than will upper nodes. The difference might be 5-8 degrees celcius. You may reduce the diffrerence by playing with the fan speeds of bottom and top nodes. If temperature builds-up on CPU or GPU, than the computer will automatically reduce the clock speed to prevent components to overheat. This will decrease the compute performance of the system. From outside, the system looks perfrectly operational but actually it works with reduced performance due to clock speed reduction.You may check if the temperature is at the limit or if there is any temperature difference between the bottom node and the upper node by using several monitoring programs. Riva Tuner, CPU-Z and GPU-Z are among them.

    Good luck with the rest of the setup.

  2. Thanks for Your insight. I indeed will follow temperatures, currently I’m using Everest Ultimate and GPU-Z for that. If temperatures will rise too high, I will modify cooling. Until that I feel quite safe.


  3. The first idea looking at the schematics was, assuming free flow after the cold air fan and local only processor cooling was that according to Bernoulli law one has to have the “tube” wider in the beginning of the distribution sequence (Node 4) and narrower at the end of it (Node 1).

    Mst gives a solution to this by suggesting different fan speeds for each level node of GPUs and I guess he is correct. This is a fair compensation for the lack of the narrowing distribution tube.

    As the load distribution and min-max overall load level on the nodes is unknown then for extra high availability I would suggest testing the whole thingy by creating artificial load on the nodes for testing purposes, before making any final conclusions. Say by loading all GPU-s or only the top GPU by, say, with calculation of nth position of pi or similar. High end GPU-s have a significant thermal emission rise at higher loads and there may be some not so obvious flow and temperature distribution effects in the game.

    One devilish detail is that while it is pretty fair to assume that from the upper GPU sections, hot air will travel to AC then depending on the actual physical layout, with the lowest section there may rise some issues, especially at higher load situations. Cold air inlet has 4 times less the cross-section compared to the hot air outlet. If the inlet and lowest hot air outlet are too close to each other (and especially if the “floor” is too close to the inlet) then the differential in-out air flows will create significant turbulence resulting in hot air to be recirculated in the system reducing overall cooling efficiency. Some simple spoiler between inlet and outlet 4 will do the trick.

    If you want to get really scientific (for the good of the community), you have two rather cheap technologies to get the ultimate truth with (
    – industrial thermal cams that you can rent for a couple of thousand EEK with a specialist operator for several hours. This will show you the true temperature distribution
    – smoke generator (similar to the one used in discotheques but meant for testing air-tightness of buildings). This will show you the actual airflows at high speeds where you can forget about laminar flow, but turbulence, even when using Node level sucking fans can create magic (negative) effects.

  4. Very informal contribution by alanrein.

    Hot air coming back from the cold air inlet can be a big problem. May be jaak already finished the set-up of this system now, but for the completeness of this discussion, I would like to suggest the below configuration ;

    * The hot air outlets are oriented to face to a window. The cold air inlet might be in opposite side (right bottom in the above drawing).

    * A window fan (such as the one below) is placed to the window

    * By means of some air blockage, hot air outlets are isolated from the rest of the room and the hot air is directed to the window fan.

    This configuration will have 3 benefits,
    1. The room temperature will not rise much nor we worry about the hot air coming back, so the cold air inlet will surely be getting cold air in.
    2. The air conditioning of the room will be much reduced, if needed at all.
    3. The noise generated by the hot air outlet fans will be suppressed by the air blockage material. Especially, the GTX295’s, if fully loaded, may contrubute to the overall noise a lot. They are at the hot air outlet side in this system, I believe.

  5. I meant informative, not informal regarding to the commends of alanrain. Sorry for my bad English.

Comments RSS TrackBack Identifier URI

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s