DGX A100 System Firmware Update Container Version 20.11.3
The DGX Firmware Update container version 20.11.3 is available.
Package name:
nvfw-dgxa100_20.11.3_201124.tar.gzRun file name:
nvfw-dgxa100_20.11.3_201124.runImage name:
nvfw-dgxa100:20.11.3
Highlights and Changes in this Release
This release is supported with the following DGX OS software -
DGX OS 4.99.11
DGX OS 5.0 or later
Includes firmware updates to resolve PCIe advanced error reporting (AER) issues.
The BMC update includes software security enhancements.
See the NVIDIA Security Bulletin 5010 for details.
Changed the policy for updating the active/inactive BMC images. See Updating Components with Secondary Images.
Removed the need to manually stop certain services before updating on DGX OS 5.0.
See DGX A100 System Firmware Changes for the list of changes in individual components.
Contents of the DGX A100 System Firmware Container
This container includes the firmware binaries and update utilities for the firmware listed in the following table. The update time for each component is provided for reference. Total update time if all components are updated is approximately 2 hours and 20 minutes.
Component |
Version |
Key Changes |
Update Time |
|---|---|---|---|
BMC (via CEC) |
00.13.16 |
The BMC update includes software security enhancements. See the NVIDIA Security Bulletin 5010 for details. |
25 minutes |
SBIOS |
0.30 |
7 minutes |
|
Broadcom 88096 PCIe switch board |
0.1.8 |
Updated preset values to address PCIe advanced error reporting (AER) issues. |
8 minutes |
BMC CEC SPI |
v3.25 |
Improved BMC update time and reliability. |
8 minutes |
PEX88064 Retimer |
0.F.0 |
Improved error handling of downstream switches. |
7 minutes |
PEX88080 Retimer |
0.F.0 |
Improved error handling of downstream switches. |
7 minutes |
NvSwitch BIOS |
92.10.14.00.01 |
Added support for a new out-of-band SMBPBI query to retrieve FUB revocation status. |
8 minutes |
VBIOS (A100 40GB) |
92.00.19.00.10 |
Improved VBIOS compatibility. |
7 minutes |
VBIOS (A100 80GB) |
92.00.36.00.01 |
New addition to the container. |
Same as above. |
U.2 NVMe (Samsung) |
EPK99B5Q |
Enabled Relaxed Ordering. |
6 minutes |
FPGA (GPU sled) |
2.9c |
Added to the container. Implements miscellaneous bug fixes. |
40 minutes |
CEC1712 SPI (GPU sled) |
3.5 |
Added to the container. Improved update time and reliability. |
7 minutes |
Updating Components with Secondary Images
Some firmware components provide a secondary image as backup. The following is the policy when updating those components:
SBIOS: The two images are referred to as active and inactive, where the active is the currently running image and the inactive is the backup image. When using
update_fw all, the update container updates both active and inactive images.BMC: The two images are referred to as active and inactive, where the active is the currently running image and the inactive is the backup image. The update container can only update the inactive image, and will update it only if the active image needs to be updated. After the update is completed, the updated inactive image becomes the active image. Because the active image is now updated, subsequent
update_fw allcommands will not update the inactive image. To update the inactive image in this case, useupdate_fw BMC --inactive --force.
Instructions for Updating Firmware
This section provides a simple way to update the firmware on the system using the firmware update container. It includes instructions for performing a transitional update for systems that require it. The commands use the .run file, but you can also use any method described in Using the DGX A100 FW Update Utility.
Caution
Stop all unnecessary system activities before attempting to update firmware, and do not add additional loads on the system (such as Kubernetes jobs or other user jobs or diagnostics) while an update is in progress. A high GPU workload can disrupt the firmware update process and result in an unusable component.
Check if updates are needed by checking the installed versions.
$ sudo nvfw-dgxa100_20.11.3_201124.run show_version
If there is “no” in any up-to-date column for updatable firmware, then continue with the next step.
If all up-to-date column entries are “yes”, then no updates are needed and no further action is necessary.
Perform the update for all firmware supported by the container.
$ sudo nvfw-dgxa100_20.11.3_201124.run update_fw all
Depending on the firmware that is updated, you may be prompted to either reboot the system or power cycle the system.
If you are prompted to reboot, issue
$ sudo reboot
If you are prompted to power cycle, you can issue the following two commands (there is no output with the first command).
$ sudo ipmitool raw 0x3c 0x04 $ sudo ipmitool chassis power cycle
After rebooting or power cycling the system, you may need to perform another
update_fw allto update other firmware.Either repeat Step 1 to check if updates are needed and then perform Step 2 if needed, or
Repeat Step 2 just in case updates are needed.
If you perform another
update_fw all, you may be prompted again to either reboot the system or power cycle the system.See DGX A100 Firmware Update Process for more information about the update process.
Rename the firmware update log file (the update generates
/var/log/nvidia-fw.logwhich you should rename).Example:
$ sudo mv /var/log/nvidia-fw.log /var/log/nvidia-fw-large.log
Refer to Firmware Update Log File Size Impacts nvsm dump health for more information.
You can verify the update by issuing the following.
$ sudo nvfw-dgxa100_20.11.3_201124.run show_version
Example output for a DGX A100 320GB system
BMC DGX
=========
Image Id Status Location Onboard Version Manifest up_to_date
0:Active Boot Online Local 00.13.16 00.13.16 yes
1:Inactive Updatable Local 00.13.16 00.13.16 yes
CEC
============
Onboard Version Manifest up-to-date
MB_CEC(enabled) 3.25 3.25 yes
SBIOS
=======
Image Id Method Onboard Version Manifest up_to_date
0:Inactive Updatabl afulnx 0.30 0.30 yes
1:Active Boot 0.30 0.30 yes
Video BIOS
============
Bus Model Onboard Version Manifest up-to-date
0000:07:00.0 A100-SXM4-40GB 92.00.19.00.10 92.00.19.00.10 yes
0000:0f:00.0 A100-SXM4-40GB 92.00.19.00.10 92.00.19.00.10 yes
0000:47:00.0 A100-SXM4-40GB 92.00.19.00.10 92.00.19.00.10 yes
0000:4e:00.0 A100-SXM4-40GB 92.00.19.00.10 92.00.19.00.10 yes
0000:87:00.0 A100-SXM4-40GB 92.00.19.00.10 92.00.19.00.10 yes
0000:90:00.0 A100-SXM4-40GB 92.00.19.00.10 92.00.19.00.10 yes
0000:b7:00.0 A100-SXM4-40GB 92.00.19.00.10 92.00.19.00.10 yes
0000:bd:00.0 A100-SXM4-40GB 92.00.19.00.10 92.00.19.00.10 yes
Switches
============
PCI Bus# Model Onboard Version Manifest up-to-date
DGX - 0000:91:00.0(U261) 88064_Retimer 0.F.0 0.F.0 yes
DGX - 0000:88:00.0(U260) 88064_Retimer 0.F.0 0.F.0 yes
DGX - 0000:4f:00.0(U262) 88064_Retimer 0.F.0 0.F.0 yes
DGX - 0000:48:00.0(U225) 88080_Retimer 0.F.0 0.F.0 yes
DGX - 0000:c4:00.0 LR10 92.10.14.00.01 92.10.14.00.01 yes
DGX - 0000:c5:00.0 LR10 92.10.14.00.01 92.10.14.00.01 yes
DGX - 0000:c2:00.0 LR10 92.10.14.00.01 92.10.14.00.01 yes
DGX - 0000:c6:00.0 LR10 92.10.14.00.01 92.10.14.00.01 yes
DGX - 0000:c3:00.0 LR10 92.10.14.00.01 92.10.14.00.01 yes
DGX - 0000:c7:00.0 LR10 92.10.14.00.01 92.10.14.00.01 yes
DGX - 0000:01:00.0(U1) PEX88096 1.8 1.8 yes
DGX - 0000:81:00.0(U3) PEX88096 1.8 1.8 yes
DGX - 0000:41:00.0(U2) PEX88096 1.8 1.8 yes
DGX - 0000:b1:00.0(U4) PEX88096 1.8 1.8 yes
Known Issues
Firmware Update Log File Size Impacts nvsm dump health
Issue
After running the container, the generated log file (/var/log/nvidia-fw.log) can grow to up to tens of gigabytes in size, depending on the firmware that gets updated. If, at a later time, you run nvsm dump health, the command might time out and fail if the file size is too large.
Explanation
To avoid problems running nvsm dump health, rename the generated firmware update log file after updating the firmware.
Example:
$ sudo mv /var/log/nvidia-fw.log /var/log/nvidia-fw-large.log
BMC UI May not be Accessible from Mac OS
Issue
When attempting to connect to the DGX A100 BMC from a system with Mac OS, a “Your connection is not private” message appears that prevents access to the BMC.
Explanation
Starting with version 0.13.6, the BMC provides a self-signed certificate which Mac OS flags in the browser. Most browsers will let you either accept the risk and continue, or add the certificate to the keychain and continue. The Chrome and Opera browsers, however, do not provide these options and so Mac OS users will not be able to access the BMC from the Chrome or Opera browser.
To access the DGX A100 BMC, Mac OS users can use Safari or Firefox, which provide an access path.
Update Timeout Reported for Motherboard CEC
Issue
The update progress output reports ” Update_timeout ” for the motherboard CEC (MB_CEC) when using the .run file without Docker installed.**Example**
+----------------------------------------------------------------------------------+
|+--------------------------------------------------------------------------------+|
|| !!!!! Firmware Update In Progress !!!!! ||
|| Status: reflash BMC firmware ||
|+--------------------------------------------------------------------------------+|
| Onboard Manifest Update Status |
|MB_CEC 3.05 3.25 Update timeout |
Explanation
This message can be ignored provided that the MB_CEC update is successful.
Example success message:
Success:
Installed firmware 3.25 on MB_CEC
Boot Order in the SBIOS Reverts to the Default
Issue
After updating the SBIOS, any changes in the boot order that you have made are not preserved and the boot order specified in the SBIOS reverts to the following.
Boot Option #1 [Hard Disk]
Boot Option #2 [NVME]
Boot Option #3 [USB CD/DVD]
Boot Option #4 [USB Hard Disk]
Boot Option #5 [USB Key]
Boot Option #6 [Network]
Explanation
You will need to prepare for the change when restarting the system, and then enter the SBIOS setup to edit the boot order as needed.