Remove a GPU (trained technician only)

Use this information to remove a GPU.

About this task

Required tools

Make sure you have the required tools listed below in hand to properly replace the component.

• SD665-N V3 Water Loop Service Kit (The water loop carrier in the Service Kit is reusable, it is recommended to keep it at the facility where the server operates for future replacement needs.)

• SD665-N V3 Water Loop Putty Pad Kit

• SD665-N V3 SXM5 PCM Fixture

• SXM5 PCM Kit (for removing PCM from GPU)

  Putty pad cannot be reused. Whenever the water loop is removed, putty pads must be replaced with new ones before reinstalling the water loop.

• Screws and screwdrivers

  Prepare the following screwdrivers to ensure you can install and remove corresponding screws properly.

  Screwdriver Type	Screw Type
  Hex screwdriver	6 mm hex head screwdriver
  Torx T10 head screwdriver	Torx T10 screw
  Phillips #1 head screwdriver	Phillips #1 screw
  Phillips #2 head screwdriver	Phillips #2 screw

Attention

• Read Installation Guidelines and Safety inspection checklist to ensure that you work safely.

• Turn off the corresponding DWC tray that you are going to perform the task on.

• Disconnect all external cables from the enclosure.

• Use extra force to disconnect QSFP cables if they are connected to the solution.

• To avoid damaging the water loop, always use the water loop carrier when removing, installing or folding the water loop.

• A torque screwdriver is available for request if you do not have one at hand.

The following illustration shows the GPU numbering.

Figure 1. GPU numbering

Procedure

1. Make preparations for this task.
   1. Remove the tray from the enclosure. See Remove a DWC tray from the enclosure.
   2. Remove the tray cover. See Remove the tray cover.
   3. Remove the cross braces. See Remove the cross braces.
   4. Remove the bus bar. See Remove the bus bar.
   5. Remove the MCIO cables. Follow the guidance and routing information in Internal cable routing.
2. Remove the carrier board power cable from the GPU node.
   Figure 2. Carrier board power cable removal
   
   

   Cable	From (carrier board)	To (GPU node power distribution board)
   1 Carrier board power cable	Power and side band connector	Power connector

3. Remove cable tie from the GPU board.
   Figure 3. Cable tie removal
   
   
4. Remove the Hex screw (x1) and the PH1 screws (x3) from the water loop with a torque screwdriver set to the proper torque.
   Note

   For reference, the torque required for the screws to be fully tightened/removed is 5.0+/- 0.5 lbf-in, 0.55+/- 0.05 N-M.

   Figure 4. Water loop Hex and PH1 screws removal (GPU node)
   
   
5. Remove water loop screws and quick connect screws (x13 Torx T10 screws) with a torque screwdriver set to the proper torque.
   Note

   For reference, the torque required for the screws to be fully tightened/removed is 5.0+/- 0.5 lbf-in, 0.55+/- 0.05 N-M.

   Figure 5. Water loop Torx T10 screws removal (GPU node)
   
   
6. Remove the quick connect screws (x4 Torx T10) with a torque screwdriver set to the proper torque.
   Note

   For reference, the torque required for the screws to be fully tightened/removed is 5.0+/- 0.5 lbf-in, 0.55+/- 0.05 N-M.

   Figure 6. Quick connect screw removal (GPU node)
   
   
7. Follow the screw removal sequence specified on the network board label, and remove network cold plate screws (x8 Torx T10 screws) with a torque screwdriver set to the proper torque.
   Note

   For reference, the torque required for the screws to be fully tightened/removed is 5.0+/- 0.5 lbf-in, 0.55+/- 0.05 N-M.

   Figure 7. Network card screw removal
   
   
8. Loosen GPU cold plate screws (x16 Torx T10 screws) in the diagonal pattern with a torque screwdriver set to the proper torque.
   Note

   For reference, the torque required for the screws to be fully tightened/loosened is 3.5 +/- 0.5 lb-In (0.4 +/ -0.06 N-m).

   Figure 8. Loosening GPU cold plate screw
   
   
9. Release the front and rear MISC conduction plates of the water loop from the GPU board.
   1. Insert a flat head screwdriver into the gaps between the MISC conduction plates (front and rear) and the GPU board. Then, slightly rotate the flat head screwdriver.
      Note

      Locations of the gaps for inserting flat head screwdriver is shown in the illustration below.
   2. The front and rear MISC conduction plates release from the GPU board slightly.
      Figure 9. Releasing the front and rear MISC conduction plates release from the GPU board
      
      
10. Release the GPU cold plates from the GPUs.
    1. There are notches on the sides of the GPU cold plates for inserting a flat head screwdriver. The locations of the notches are shown in the illustration below.
       Attention

       Inserting the flat head screwdriver ONLY to the notches circled in the illustration below. Otherwise, the screwdriver may damage the GPUs.

       Figure 10. Notches for releasing GPU cold plates
       
       
    2. Insert a flat screwdriver into all the notches shown in the illustration; then, slightly rotate the screwdriver to release the GPU cold plates from the GPUs.
       Figure 11. Releasing the GPU cold plate from the GPU
       
       
11. Gently put the water loop carrier down onto the water loop and ensure it is seated firmly on the water loop.
    Figure 12. Water loop carrier installation (GPU node)
    
    
12. Tighten water loop carrier screws (x20 Phillips #2 screws) with a torque screwdriver set to the proper torque.
    Note

    For reference, the torque required for the screws to be fully tightened/removed is 5.0+/- 0.5 lbf-in, 0.55+/- 0.05 N-M.

    Figure 13. Water loop screw and quick connect screw removal (GPU node)
    
    
13. Carefully rotate the GPU node side water loop so that it is sitting on top of the Compute node side water loop.
    Figure 14. Folding the GPU node side water loop onto the Compute node side water loop
    
    
14. Immediately clean the PCM off all the GPUs with alcohol cleaning pads. Gently clean the PCM to avoid GPU damages.
    Attention

    • It is recommended to clean the PCM while it is in liquid state.

    • The electrical components around the die on the GPUs are extremely delicate. When removing the PCM and cleaning the GPU die, avoid touching the electrical components to prevent damages.

    Figure 15. Cleaning PCM off from all GPUs
    
    
15. With alcohol cleaning pads, wipe off any remaining putty pad and PCMs from the water loop and components in the GPU node.
    Figure 16. Cleaning putty pads from water loop
    
    
16. Locate the GPU to be removed. Then, remove the four Torx T15 screws in the sequence shown in the illustration below with a torque screwdriver set to the proper torque. Then, carefully remove the GPU out of the GPU board.
    Note

    For reference, the torque required for the screws to be fully tightened/removed is 0.45-0.56 N-m, 4.0-5.0 in-lbf.

    Figure 17. GPU removal

    Screw installation sequence: → → →