Maintaining and Servicing the NVIDIA DGX-1

Be sure to familiarize yourself with the NVIDIA Terms & Conditions documents before attempting to perform any modification or repair to the DGX-1. These Terms & Conditions for the DGX-1 can be found through the NVIDIA DGX Systems Support page.

Problem Resolution and Customer Care

Log on to the NVIDIA Enterprise Services site for assistance with troubleshooting, diagnostics, or to report problems with your DGX-1.

Refer to Submitting BMC Log Files for instructions on how to obtain the BMC log files to assist in troubleshooting.

Restoring the DGX-1 Software Image

If the DGX-1 software image becomes corrupted or the OS SSD was replaced after a failure, restore the DGX-1 software image to its original factory condition from a pristine copy of the image.

The process for restoring the DGX-1 software image is as follows:

  1. Obtain an ISO file that contains the image from NVIDIA Support Enterprise Services as explained in Obtaining the DGX-1 Software ISO Image and Checksum File.
  2. Restore the DGX-1 software image from this file either remotely through the BMC or locally from a bootable USB flash drive.

Obtaining the DGX-1 Software ISO Image and Checksum File

To ensure that you restore the current version of the DGX-1 software image, obtain the correct ISO image file from NVIDIA Support Enterprise Services. A checksum file is provided for the image to enable you to verify the bootable installation medium that you create from the image file.
  1. Log on to the NVIDIA Enterprise Services site.
  2. Click the Announcements tab to locate the download links for the DGX-1 software image.
  3. Download the ISO image and its checksum file and save them to your local disk. The ISO image is also available in an archive file. If you download the archive file, be sure to extract the ISO image before proceeding.

Re-Imaging the System Remotely

These instructions describe how to re-image the system remotely through the BMC. For information about how to restore the system locally, see Re-Imaging the System From a USB Flash Drive.

Before re-imaging the system remotely, ensure that the correct DGX-1 software image is saved to your local disk. For more information, see Obtaining the DGX-1 Software ISO Image and Checksum File.
  1. Connect to the BMC and change user privileges.
    1. Open a Java-enabled web browser within your LAN and go to http://IPMI-IP-address/, then log in.

      Use Firefox or Internet Explorer. Google Chrome is not officially supported by the BMC.

    2. From the top menu, click Configuration and then select User Management.
    3. Select the user name that you created for the BMC, then click Modify User.
    4. In the Modify User dialog, select the VMedia check box to add it to the extended privileges for the user, then click Modify.



  2. Set up the ISO image as virtual media.
    1. From the top menu, click Remote Control and select Console Redirection.



    2. Click Java Console to open the remote JViewer window. Make sure pop-up blockers are disabled for this site.
    3. From the JViewer top menu bar, click Media and then select Virtual Media Wizard.





    4. From the CD/DVD Media: I section of the Virtual Media dialog, click Browse and then locate the re-image ISO file and click Open.

      You can ignore the device redirection warning at the bottom of the Virtual Media wizard as it does not affect the ability to re-image the system.

    5. Click Connect CD/DVD, then click OK at the Information dialog.

      The Virtual Media window shows that the ISO image is connected.





    6. Close the window.

      The CD ROM icon in the menu bar turns green to indicate that the ISO image is attached.





  3. Reboot, install the image, and complete the DGX-1 setup.
    1. From the top menu, click Power and then select Reset Server.



    2. Click Yes and then OK at the Power Control dialogs, then wait for the system to power down and then come back online.
    3. At the boot selection screen, select Install DGX-1 OS and then press [Enter].

      If you are an advanced user who is not using the RAID disks as cache and want to keep data on the RAID disks, then select Install DGX Server without formatting RAID. See the section Retaining the RAID Partition While Installing the OS for more information.

      The DGX-1 will reboot from CDROM0 1.00, and proceed to install the image. This can take approximately 15 minutes.

      Note: The Mellanox InfiniBand driver installation may take up to 10 minutes.

      After the installation is completed, the system ejects the virtual CD and then reboots into the OS.

Refer to Setting Up the DGX-1 for the steps to take when booting up the DGX-1 for the first time after a fresh installation.

Creating a Bootable Installation Medium

After obtaining an ISO file that contains the software image from NVIDIA Support Enterprise Services, create a bootable installation medium, such as a USB flash drive or DVD-ROM, that contains the image.

Note: If you are restoring the software image remotely through the BMC, you do not need a bootable installation medium and you can omit this task.

Creating a Bootable USB Flash Drive by Using the dd Command

On a Linux system, you can use the dd command to create a bootable USB flash drive that contains the DGX-1 software image.

Note: Because the image is a hybrid ISO image, you must convert and copy the image to perform a device bit copy of the image. You cannot perform a simple file copy of the image.

Ensure that the following prerequisites are met:

  • The correct DGX-1 software image is saved to your local disk. For more information, see Obtaining the DGX-1 Software ISO Image and Checksum File.
  • The USB flash drive meets these requirements:
    • The USB flash drive has a capacity of at least 4 GB.
    • The partition scheme on the USB flash drive is a GPT partition scheme for UEFI.
  1. Plug the USB flash drive into one of the USB ports of your Linux system.
  2. Obtain the device name of the USB flash drive by running the lsblk command.
    lsblk

    You can identify the USB flash drive from its size, which is much smaller than the size of the SSDs in the DGX-1, and from the mount points of any partitions on the drive, which are under /media.

    In the following example, the device name of the USB flash drive is sde.

    ~$ lsblk
    NAME   MAJ:MIN RM   SIZE RO TYPE MOUNTPOINT
    sda      8:0    0   1.8T  0 disk 
    |_sda1   8:1    0   121M  0 part /boot/efi
    |_sda2   8:2    0   1.8T  0 part /
    sdb      8:16   0   1.8T  0 disk 
    |_sdb1   8:17   0   1.8T  0 part 
    sdc      8:32   0   1.8T  0 disk 
    sdd      8:48   0   1.8T  0 disk 
    sde      8:64   1   7.6G  0 disk 
    |_sde1   8:65   1   7.6G  0 part /media/deeplearner/DGXSTATION
    ~$
  3. As root, convert and copy the image to the USB flash drive.
    sudo dd if=path-to-software-image bs=2048 of=usb-drive-device-name
    CAUTION:
    The dd command erases all data on the device that you specify in the of option of the command. To avoid losing data, ensure that you specify the correct path to the USB flash drive.

Creating a Bootable USB Flash Drive by Using Akeo Rufus

On a Windows system, you can use the Akeo Reliable USB Formatting Utility (Rufus) to create a bootable USB flash drive that contains the DGX-1 software image.

Ensure that the following prerequisites are met:

  1. Plug the USB flash drive into one of the USB ports of your Windows system.
  2. Download and launch the Akeo Reliable USB Formatting Utility (Rufus).



  3. Under Partition scheme and target system type, select GPT partition scheme for UEFI.
  4. Select the Create a bootable disk using option and from the dropdown menu, select ISO image.
  5. Click the optical drive icon and open the DGX-1 software ISO image.
  6. Click Start. Because the image is a hybrid ISO file, you are prompted to select whether to write the image in ISO Image (file copy) mode or DD Image (disk image) mode.



  7. Select Write in ISO Image mode and click OK.

Re-Imaging the System From a USB Flash Drive

These instructions describe how to re-image the system from a USB flash drive. For information about how to restore the system remotely, see Re-Imaging the System Remotely.

Before re-imaging the system from a USB flash drive, ensure that you have a bootable USB flash drive that contains the current DGX-1 software image.

  1. Plug the USB flash drive containing the OS image into the DGX-1.
  2. Connect a monitor and keyboard directly to the DGX-1.
  3. Boot the system and press F11 when the NVIDIA logo appears to get to the boot menu.
  4. Select the USB volume name that corresponds to the inserted USB flash drive, and boot the system from it.
  5. When the system boots up, select Install DGX-1 OS on the startup screen and then press Enter.

    If you are an advanced user who is not using the RAID disks as cache and want to keep data on the RAID disks, then select Install DGX Server without formatting RAID. See the section Retaining the RAID Partition While Installing the OS for more information.

The DGX-1 will reboot and proceed to install the image. This can take more than 15 minutes.

Note: The Mellanox InfiniBand driver installation may take up to 10 minutes.

After the installation is completed, the system then reboots into the OS.

Refer to Setting Up the DGX-1 for the steps to take when booting up the DGX-1 for the first time after a fresh installation.

Retaining the RAID Partition While Installing the OS

This information describes an installation option that is available starting with DGX OS Server 3.1.1.

The re-imaging process creates a fresh installation of the DGX OS. During the OS installation or re-image process, you are presented with a boot menu when booting the installer image. The default selection is Install DGX Software. The installation process then repartitions all the SSDs, including the OS SSD as well as the RAID SSDs, and the RAID array is mounted as /raid. This overwrites any data or file systems that may exist on the OS disk as well as the RAID disks.

Since the RAID array on the DGX-1 is intended to be used as a cache and not for long-term data storage, this should not be disruptive. However, if you are an advanced user and have set up the disks for a non-cache purpose and want to keep the data on those drives, then select the Install DGX Server without formatting RAID option at the boot menu during the boot installation. This option retains data on the RAID disks and performs the following:

  • Installs the cache daemon but leaves it disabled by commenting out the RUN=yes line in /etc/default/cachefilesd.
  • Creates a /raid directory, leaves it out of the file system table by commenting out the entry containing “/raid” in /etc/fstab.
  • Does not format the RAID disks.
When the installation is completed, you can repeat any configurations steps that you had performed to use the RAID disks as other than cache disks.
You can always choose to use the RAID disks as cache disks at a later time by enabling cachefilesd and adding /raid to the file system table as follows:
  1. Uncomment the #RUN=yes line in /etc/default/cachefiled.
  2. Uncomment the /raid line in etc/fstab.
  3. Run the following:
    1. Mount /raid.
      sudo mount /raid
    2. Reload the systemd manager configuration.
      systemctl daemon-reload
    3. Starty the cache daemon.
      sytemctl start cachefilesd.server

These changes are preserved across system reboots.

Updating the System BIOS

You can update the system BIOS remotely through the BMC. Before updating the system BIOS, the system must be turned off through the BMC according to the instructions in this section.

  1. Obtain the BIOS image.
    1. Log on to NVIDIA Enterprise Services and click the Announcements tab to locate the DGX-1 software image archive.
    2. Download the image archive and then extract the .bin file.
  2. Log on to the BMC and shut down the DGX-1.
    1. Open a Java-enabled web browser within your LAN and go to http:\\<IPMI IP address>\, then log in.

      Use Firefox or Internet Explorer. Google Chrome is not officially supported by the BMC.

    2. From the top menu, click Remote Control and then select Server Power Control.
    3. At the Power Control and Status screen, select the Power Off Server - Orderly Shutdown option, then click Perform Action.

      You can verify that the DGX-1 is shut down by noting that all the Power Control and Status options are grayed out except for the Power On Server option.

  3. Update the system BIOS.
    1. From the top menu, click Firmware Update, select BIOS Update, and then click Enter Update Mode.

    2. Click OK at the Are you sure to enter update mode? dialog.
    3. From the BIOS Upload screen, click Browse at the Select Firmware to Upload step, then navigate the explorer windows to locate the file you downloaded and select it.

    4. Be sure all the check boxes under Select Preserve Configuration are cleared. This ensures that the BIOS reverts to its fail-safe default settings for a reliable update.
    5. Click Upload Firmware to start the process of installing the updated BIOS.

      You are asked to wait while the image is verified.

    6. Click OK at the Proceed? dialog to start the actual upgrade process.

      The BIOS Flash Status screen shows the upgrade progress, which should take a couple of minutes to complete.

      Note: Do not interrupt the upgrade process once it has started.
  4. After the upgrade process has completed, you can use the top menu to turn the system back on.
    1. From the top menu, click Remote Control and then select Server Power Control.
    2. Select the Power On Server option, and then click Perform Action.
  5. To verify that the BIOS was updated with the proper file, press [F2] or [Del] to enter the BIOS setup screen when the system reboots, then compare the Project Version with the update filename.

Updating the BMC

You can update the BMC remotely using the IPMI port, This can be done while the system is powered on and with applications running.

  1. Obtain the BMC image.
    1. Log on to NVIDIA Enterprise Services and click the Announcements tab to locate the DGX-1 software image archive.
    2. Download the image file.
  2. Open a Java-enabled web browser within your LAN and go to http:\\<IPMI IP address>\, then log in to the BMC.

    Use Firefox or Internet Explorer. Google Chrome is not officially supported by the BMC.

  3. If you’re using DHCP and choose not to preserve the network configuration, then obtain the MAC address for the BMC.

    If the BMC is connected to a network via DHCP, the IP address could change after the update. Follow these substeps to obtain the MAC address in order to connect to the BMC after the update, in case the IP address changes. You can skip these steps if a static IP is used.

    1. From the top menu, click Configuration and then select Network.
    2. Note the MAC address.
  4. From the top menu, click Firmware Update and then select Firmware Update from the drop-down menu.
  5. Click Enter Preserve Configuration, then set the IPMI Preserve Status to Preserve and all others to Overwrite.

    Note: Be sure to set IPMI to Preserve in order to preserve your BMC login credentials. If you fail to do this, the BMC username/password will be set to qct.admin/qct.admin. If this happens, then be sure to enter the BMC dashboard and go to Configuration->Users to add a new user account and disable the qct.admin account after updating the BMC.
  6. If necessary, click Firmware Update again from the top menu and then select Firmware Update from the drop-down menu to return to the Firmware Update page.
  7. Click Enter Update Mode, then click OK at the confirmation dialog.

    After entering Update Mode, aborting the operation or even resizing the browser windows will terminate the session and reset the BMC. If this happens, you will need to close and then reopen the browser to initiate a new session. You may need to wait several minutes for the BMC to reset.

  8. At the Upload Firmware prompt, click Browse to locate and select the firmware image file.

    Select the encrypted file (the file with the "_enc" suffix on the file extension), as the BMC requires the firmware image to be encrypted.

  9. Click Upload to transfer the image to the BMC.
  10. At the Select Based Firmware Update prompt. select Full Flash and then click Proceed.

  • When the BMC firmware update is completed, the BMC resets and the remote session terminates.
  • To initiate a new BMC session, close and then reopen the browser.
  • The BMC can take as much as 10 minutes to reset itself. During this time, the BMC will be unresponsive.

Replacing the System and Components

Be sure to familiarize yourself with the NVIDIA Terms & Conditions documents before attempting to perform any modification or repair to the DGX-1. These Terms & Conditions for the DGX-1 can be found through the NVIDIA DGX Systems Support page.

Contact NVIDIA Enterprise Customer support to obtain an RMA number for any system or component that needs to be returned for repair or replacement.

The following components are customer-replaceable:
  • Solid State Drives (SSDs)
  • Power Supplies
  • Fan Modules
  • DIMMs

Return the failed components to NVIDIA. Low-cost items such as power supplies and fans do not need to be returned.

Replacing the System

When returning a DGX-1 under RMA, consider the following points.

SSDs

If necessary, you can remove and keep the SSDs prior to shipping the system back for replacement. If you already received a replacement system and you want to keep the original SSDs, install the new SSDs into the defective system when shipping it back.

Bezel

Be sure to include the bezel when returning the system.

Replacing an SSD

Access the SSDs from the front of the DGX-1. You can hot swap the SSDs as follows:

  1. If not already removed, remove the bezel by grasping the bezel by the side handles and then pulling the bezel straight off the front of the DGX-1.
    Note:CAUTION: Be careful not to accidentally press the power button that is on the right edge of the DGX-1 when removing or installing the bezel.
  2. Locate the SDD that you want to replace, then press the round button at the top edge to release the latch.
  3. Pull the latch down and then out to unseat the SSD assembly.



  4. Continue pulling the SSD assembly to completely remove it from the unit.
  5. Using a phillips screwdriver, remove the four screws attaching the SSD to the hot-swap tray.



  6. Save the screws for the replacement.
  7. Mount the replacement SSD to the hot-swap tray using the four screws. Make sure that the connector is on the open edge side of the tray.
  8. With the round button at the top, insert the assembly into the appropriate bay, then push the assembly all the way in.
  9. Press the latch against the assembly to completely seat the assembly.
  10. Reattach the bezel.

    With the bezel positioned so that the NVIDIA logo is visible from the front and is on the left-hand side, line up the pins near the corners of the DGX-1 with the holes in back of the bezel, then gently press the bezel against the DGX-1. The bezel is held in place magnetically.

    Note:CAUTION: Be careful not to accidentally press the power button that is on the right edge of the DGX-1 when removing or installing the bezel.

Recreating the Virtual Drives

After you have replaced the OS SSD, with or without any of the cache SSDs, you need to recreate the virtual drives and then re-image the system in order to recreate the partitions on all the virtual drives.

The following is an overview of the process:

  1. Clear the drive group configuration
  2. Recreate the OS Virtual Drive
  3. Recreate the Cache Virtual Drive
  4. Re-image the System

These instructions apply only if you have replaced the OS SSD, with or without one or more of the cache SSDs. If you have replaced only one or more of the cache SSDs, and not the OS SSD, then follow the instructions in the section Recreating the RAID 0 Array

Access the BIOS Setup Utility

RAID configuration is accomplished through the BIOS setup utility.

  1. Connect a display (1024x768 or lower resolution) and keyboard to the DGX-1.
  2. Turn the DGX-1 on or reboot.
  3. At the NVIDIA logo boot screen, press [F2] or [Del] to enter the BIOS setup screen.

  4. Select the Advanced tab from the top menu and then Scroll down and select the MegaRAID Configuration Utility.

    The RAID Configuration menu appears.

If you replaced the OS drive, follow the instructions in the section Clear the Drive Group Configuration .

Clear the Drive Group Configuration

These instructions apply when you have replaced the OS drive.

  1. Select Main Menu, then select Configuration Management.

  2. Select Clear Configuration.

  3. Select Confirm [Disabled] and then select Enabled at the confirmation popup.

  4. Select Yes, then select OK at the success screen.

  5. Follow the instructions in the sections Recreate the OS Virtual Drive and then Recreate the RAID0 Virtual Drive .

Recreate the OS Virtual Drive

These instructions apply when you have replaced the OS drive. Be sure to first complete the instructions in the section Clear the Drive Group Configuration.

  1. Navigate to the RAID Utility Main Menu, then under Actions, select Configure, then select Configuration Management.

  2. Select Create Virtual Drive, then select Select Drives at the next screen. Leave all other options at their default settings as shown below.

    The list of drives under CHOOSE UNCONFIGURED DRIVES will initially be empty.
  3. To view the available drives, select Select Media Type [HDD], then change to [SSD].

  4. Under CHOOSE UNCONFIGURED DRIVES, select the 446 GB drive, then change to [Enabled] at the pop-up dialog.

  5. Confirm that only the first drive at Drive Port 0 - 3:01:00 displays as [Enabled].

  6. Scroll up and select Apply Changes.

  7. Select OK at the success screen.

    The virtual drive creation page now displays a summary of your selection. The Virtual Drive Size should be approximately 446 GB.
  8. Select Save Configuration at the top of the menu.
  9. Change the Confirm [Disabled] field to [Enabled] and then select [Yes].

  10. Select [OK] at the success screen. You have successfully re-created Virtual Drive 0, where the OS will be installed.
  11. Follow the instructions in the section Recreate the RAID0 Virtual Drive

Recreate the RAID0 Virtual Drive

These instructions apply when you have replaced the OS drive and cleared the drive group configuration.

  1. Navigate to the RAID Utility Main Menu, then under Action, select Configure, then select Configuration Management.

  2. Select Create Virtual Drive.

  3. Scroll to Select RAID Level and switch to [RAID0], if not already set.

  4. Scroll to Select Media Type and switch to [SSD].

  5. Select Select Drives.

  6. Switch all unconfigured 1TB drives to [Enabled].

  7. Select Apply Changes.

  8. Change Confirm to [Enabled], then select Yes.
  9. Select OK at the success screen. The Create Virtual Drive screen displays a summary of your selection.
  10. Verify that the summary matches your selection, then select Save Configuration.

  11. Make sure Confirm is set to [Enabled], then select Yes to confirm the change.

  12. Select OK at the success screen.
  13. Confirm and exit.
    1. Select View Drive Group Properties to confirm the configuration.

    2. Verify that your configuration screen shows that you have two virtual drives with the following properties: Virtual Drive 0 of size 446 GB (or very similar) Virtual Drive 1 of size 7 TB (or very similar).

    3. If your Drive Groups match the above, press [F10] to save these settings and reset the system.
    4. Select Save Changes and Reset, then select Yes at the confirmation prompt.

  14. Follow the instructions in the section Restoring the DGX-1 Software Image to create the partitions.

Recreating the RAID 0 Array

After replacing one of the RAID 0 cache SSDs, you need to recreate the RAID 0 array. If you replaced only the cache and not the operating system SSD, then you can use a convenient script to recreate the RAID array. The script is part of the DGX-1 software as of version 2.0.4.
To use the script, you need to get and install the StorCLI utility. For instructions, see the document Using StorCLI to Recreate the NVIDIA DGX-1 RAID 0 Array, available from the Enterprise Services site.
Note: Connect a display (1024x768 or lower resolution) and keyboard to the DGX-1 when booting the DGX-1 before recreating the RAID array. This is because the system may halt at the BIOS screen alerting you that the RAID array needs to be configured. Press C (or whichever key allows you to continue) to complete the boot process. You will be able to do this only if you are operating the DGX-1 through a direct display and keyboard connection.
  1. If you have installed the StorCLI utility, run the script by entering the following on the command line:
     $ sudo python /usr/local/bin/configure_raid_array.py -c -f 
  2. After the script has finished recreating the RAID 0 array, reboot the DGX-1 to verify that /raid is mounted and usable.
Refer to the document Using StorCLI to Recreate the NVIDIA DGX-1 RAID 0 Array for more information.

Replacing the Power Supplies

Access the power supplies from the front of the DGX-1. You can hot-swap the power supplies as follows:

  1. If not already removed, remove the bezel by grasping the bezel by the side handles and then pulling the bezel straight off the front of the DGX-1.
    Note:CAUTION: Be careful not to accidentally press the power button that is on the right edge of the DGX-1 when removing or installing the bezel.
  2. Unplug the power cord from the power connector on the fan assembly.
  3. Flip the power supply handle out.



  4. Push the green release lever to the left and simultaneously use the power supply handle to pull out the power supply.
  5. Slide the replacement power supply into the bay and push until seated.
  6. Flip the power supply handle up against the power supply.
  7. Reconnect the power cord.
    Note:IMPORTANT: Make sure that the end of the power cord cable tie is not inserted into the power supply fan. The cable tie can interfere with normal operation of the fan, resulting in failure of the power supply.
  8. Reattach the bezel. With the bezel positioned so that the NVIDIA logo is visible from the front and is on the left-hand side, line up the pins near the corners of the DGX-1 with the holes in back of the bezel, then gently press the bezel against the DGX-1. The bezel is held in place magnetically.
    Note:CAUTION: Be careful not to accidentally press the power button that is on the right edge of the DGX-1 when removing or installing the bezel.

Replacing the Fan Module

CAUTION: To avoid overheating the system, the fan module should be replaced within 25 seconds after removal.

  1. Unscrew the thumbscrews at the front of the DGX-1, then slide the DGX-1 about half way out from the rack.
  2. Squeeze together the latches at the square access openings on the top of the chassis, then flip open the top of the chassis to expose the fan modules.



  3. Squeeze the release tabs on the outer edge of the fan module you want to replace, then pull up to lift the fan module out of the unit.



  4. Replace with a new fan module using the reverse steps.

Replacing the DIMMs

Before attempting to replace any of the dual inline memory modules (DIMMs), make sure that you know the location of the faulty DIMM needing replacment. The location ID is an alpha-numeric designator, such as A0, A1, B0, B1, etc., and is reported in the BMC log files.

CAUTION: Static Sensitive Devices: - Be sure to observe best practices for electrostatic discharge (ESD) protection. This includes making sure personnel and equipment are connected to a common ground, such as by wearing a wrist strap connected to the chassis ground, and placing components on static-free work surfaces.

The DIMMs are located on the motherboard tray, which is accessible from the rear of the DGX-1.

  1. Turn off the DGX-1 and disconnect all network and power cabling.
  2. Remove the motherboard tray.
    1. Locate the locking levers for the motherboard tray at the rear of the DGX-1. There are two sets of locking levers. The locking levers for the motherboard are the bottom set.

    2. Rotate the retention clasps inward towards the center of the unit.

      The retention clasps hold the locking levers in place. Rotating the clasps inward releases the locking levers.
    3. Swing the locking levers out and then use them to pull the motherboard tray out of the unit.

      Do not pull the unit by the blue retention clasps; they may break.
    4. Set the motherboard tray on a clean work surface, and position it so that the locking levers are at the top as you look down on the tray.

      The DIMMs are on a printed circuit board on the left side of the tray.
  3. Using the figure below as a guide, locate the DIMM corresponding to the ID of the faulty DIMM as reported in the BMC log.

  4. Remove the DIMM.
    1. Press down on the side latches at both ends of the DIMM socket to push them away from the DIMM. This should unseat the DIMM from the socket.
    2. Pull the DIMM straight up to remove it from the socket.
  5. Carefully insert the replacement DIMM.
    1. Make sure the socket latches are open.
    2. Positon the DIMM over the socket, making sure that the notch on the DIMM lines up with the key in the slot, then press the DIMM down into the socket until the side latches click in place.
    3. Make sure that the latches are up and locked in place.
  6. Carefully insert the motherboard tray back into the unit, then swing the locking levers flat against the tray and secure them in place with the retention clasps.

Replacing the InfiniBand Cards

The InfiniBand cards are located on the GPU tray which is accessible from the rear of the DGX-1. Be sure you have identified the faulty InfiniBand card needing to be replaced. The slots are identified as indicated in the following image.

CAUTION: Static Sensitive Devices: - Be sure to observe best practices for electrostatic discharge (ESD) protection. This includes making sure personnel and equipment are connected to a common ground, such as by wearing a wrist strap connected to the chassis ground, and placing components on static-free work surfaces.

  1. Turn off the DGX-1 and disconnect all network and power cabling.
  2. Remove the GPU tray.
    1. Locate the locking levers for the GPU tray at the rear of the DGX-1. There are two sets of locking levers. The locking levers for the GPU tray are the top set.

    2. Rotate the retention clasps inward towards the center of the unit.

      The retention clasps hold the locking levers in place. Rotating the clasps inward releases the locking levers.
    3. Swing the locking levers out and then use then to pull the GPU tray out of the unit.

      Do not pull the unit by the blue retention clasps; they may break.
  3. Set the GPU tray on a clean work surface.
  4. At the top edge of the bracket for the InfiniBand card that you want to replace, rotate the retention clasp to free the bracket.

  5. Firmly grasp the InfiniBand card and lift it straight up out of the PCIe slot.

  6. Position the replacement InfiniBand card over the empty PCIe slot and insert it into the slot.
  7. Swing the retention clasp over the bracket to secure the bracket in place.

  8. Carefully insert the GPU tray back into the unit, then swing the locking levers flat against the tray and secure them in place with the retention clasps.
  9. Reconnect all connectors, boot the system, then perform the verification and setup steps described in the next section.

Setting Up the InfiniBand Cards

This section describes the steps needed to verify that the InfiniBand card has been replaced correctly.

  1. With the DGX-1 turned on, verify that the card was installed correctly and is recognized by the system.
    $ lspci | grep -i mellanox
    The output should show all four InfiniBand cards. Example:
     05:00.0 Infiniband controller: Mellanox Technologies MT27700 Family [ConnectX-4]
     0c:00.0 Infiniband controller: Mellanox Technologies MT27700 Family [ConnectX-4]
     84:00.0 Infiniband controller: Mellanox Technologies MT27700 Family [ConnectX-4]
     8b:00.0 Infiniband controller: Mellanox Technologies MT27700 Family [ConnectX-4] 
    If all four cards are not reported, then the card was not installed properly and should be reseated. If a card other than the officially supported Mellanox family of adapters appears, contact NVIDIA Enterprise Support.
  2. Verify that the InfiniBand drivers are present.
    $ lsmod | grep -i ib_
    The output should be a list of lb_ and mlx_ driver components. Example:
     ib_ucm 20480 0
     ib_ipoib 131072 0
     ib_cm 45056 3 rdma_cm,ib_ucm,ib_ipoib
     ib_uverbs 73728 2 ib_ucm,rdma_ucm
     ib_umad 24576 0
     mlx5_ib 192512 0
     mlx4_ib 192512 0
     ib_sa 36864 5 rdma_cm,ib_cm,mlx4_ib,rdma_ucm,ib_ipoib
     ib_mad 57344 4 ib_cm,ib_sa,mlx4_ib,ib_umad
     ib_core 143360 13 rdma_cm,ib_cm,ib_sa,iw_cm,nv_peer_mem,mlx4_ib,mlx5_ib,ib_mad,ib_ucm,ib_umad,ib_uverbs,rdma_ucm,ib_ipoib
     ib_addr 20480 3 rdma_cm,ib_core,rdma_ucm
     ib_netlink 16384 3 rdma_cm,iw_cm,ib_addr
     mlx4_core 344064 2 mlx4_en,mlx4_ib
     mlx5_core 524288 1 mlx5_ib
     mlx_compat 16384 18 rdma_cm,ib_cm,ib_sa,iw_cm,mlx4_en,mlx4_ib,mlx5_ib,ib_mad,ib_ucm,ib_netlink,ib_addr,ib_core,ib_umad,ib_uverbs,mlx4_core,mlx5_core,rdma_ucm,ib_ipoib
  3. Verify that the OFED software was installed correctly.
     $ modinfo mlx5_core | grep -i version | head -1 
    Example output:
    Version : 3.4-1.0.0
    DGX-1 OS release 1.0 should have OFED software 3.2. DGX-1 OS release 2.0 should have OFED software 3.4.
  4. Restart the InfiniBand services so that the new card is recognized.
    1. Restart the InfiniBand service.
      $ sudo service openibd restart
    2. Restart the Service Manager service.
      $ sudo service opensmd restart
    3. Verify that the service has started.
      $ service openibd status
        openibd start/running
      $ service opensmd status
        OpenSM is running...
    4. If the services do not start, verify
      • That the drivers are loaded according to step 3.
      • That the associated cables are connected to the InfiniBand ports.
      • The state of ibstat (refer to step 7)
      • Whether errors are reported in /var/log/syslog.
      If these steps do not indicate a problem and yet the services still do not start, contact NVIDIA Enterprise Support and obtain an RMA for the card.
  5. Verify the firmware version.
    $ cat /sys/class/infiniband/mlx5*/fw_ver
    Example output:
     12.17.1010
     12.17.1010
     12.17.1010
     12.17.1010
    The latest InfiniBand firmware version supported on DGX-1 OS release 1.0 is 12.16.1020, and the latest supported on release 2.0 is 12.17.1010.
  6. If you need to update the firmware, follow these steps:
    1. Initiate the firmware update.
      $ sudo /opt/mellanox/mlnx-fw-updater/mlnx_fw_updater.pl
      The script will check the firmware version of each card and update where needed. If the firmware is updated for any card, you will need to reboot the system for the changes to take effect.
    2. Reboot the system if instructed.
    3. After rebooting the system, verify that all the Mellanox InfiniBand cards are using the current firmware.
      $ cat /sys/class/infiniband/mlx5*/fw_ver
       12.17.1010
       12.17.1010
       12.17.1010
       12.17.1010
  7. Verify the physical port state for the InfiniBand cards.
    $ ibstat
    In the output text, verify that the Physical State for each card with a cable connection is LinkUp and that the port for the card is configured with a GUID. The following example output shows one card in a non-connected state, and three cards in a connected state. Relevant text is highlighted in bold.
    CA 'mlx5_0'
     CA type: MT4115
     Number of ports: 1
     Firmware version: 12.17.1010
     Hardware version: 0
     Node GUID: 0x248a0703000de288
     System image GUID: 0x248a0703000de288
     Port 1:
           State: Down
           Physical state: Polling
           Rate: 10
           Base lid: 65535
           LMC: 0
           SM lid: 0
           Capability mask: 0x2651e848
           Port GUID: 0x248a0703000de288
           Link layer: InfiniBand
    CA 'mlx5_1'
     CA type: MT4115
     Number of ports: 1
     Firmware version: 12.17.1010
     Hardware version: 0
     Node GUID: 0x248a0703000de26c
     System image GUID: 0x248a0703000de26c
     Port 1:
            State: Initializing
            Physical state: LinkUp
            Rate: 100
            Base lid: 65535
            LMC: 0
            SM lid: 0
            Capability mask: 0x2651e848
            Port GUID: 0x248a0703000de26c
            Link layer: InfiniBand
    CA 'mlx5_2'
     CA type: MT4115
     Number of ports: 1
     Firmware version: 12.17.1010
     Hardware version: 0
     Node GUID: 0x248a0703001effde
     System image GUID: 0x248a0703001effde
     Port 1:
            State: Initializing
            Physical state: LinkUp
            Rate: 100
            Base lid: 65535
            LMC: 0
            SM lid: 0
            Capability mask: 0x2651e848
            Port GUID: 0x248a0703001effde
            Link layer: InfiniBand
    CA 'mlx5_3'
     CA type: MT4115
     Number of ports: 1
     Firmware version: 12.17.1010
     Hardware version: 0
     Node GUID: 0x7cfe900300118f22
     System image GUID: 0x7cfe900300118f22
     Port 1:
             State: Initializing
             Physical state: LinkUp
             Rate: 100
             Base lid: 65535
             LMC: 0
             SM lid: 0
             Capability mask: 0x2651e848
             Port GUID: 0x7cfe900300118f22
             Link layer: InfiniBand