This utility is used for various configuration including the Confidential Computing modes of supported GPUs as well as some debug/test tasks. It is designed to be run as a privileged python3 command.
Supported CC modes are:
sudo apt install patchelf python3-pipsudo python3 ./nvidia_gpu_tools.py --gpu-name=H100 --query-cc-mode
sudo python3 ./nvidia_gpu_tools.py --gpu-name=H100 --set-cc-mode=on --reset-after-cc-mode-switch
sudo python3 ./nvidia_gpu_tools.py --gpu-bdf=45:00.0 --set-cc-mode=off --reset-after-cc-mode-switch
sudo python3 ./nvidia_gpu_tools.py --gpu-bdf=45:00.0 --debug-dump --log debug
sudo python3 ./nvidia_gpu_tools.py --gpu-bdf=45:00.0 --nvlink-debug-dump --log debug
sudo python3 nvidia_gpu_tools.py --help
NVIDIA GPU Tools version v2024.01.19o
Command line arguments: ['./nvidia_gpu_tools.py', '--help']
Usage: nvidia_gpu_tools.py [options]
Options:
-h, --help show this help message and exit
--gpu=GPU
--gpu-bdf=GPU_BDF Select a single GPU by providing a substring of the
BDF, e.g. '01:00'.
--gpu-name=GPU_NAME Select a single GPU by providing a substring of the
GPU name, e.g. 'T4'. If multiple GPUs match, the first
one will be used.
--no-gpu Do not use any of the GPUs; commands requiring one
will not work.
--log=LOG
--mmio-access-type=MMIO_ACCESS_TYPE
On Linux, specify whether to do MMIO through /dev/mem
or /sys/bus/pci/devices/.../resourceN
--recover-broken-gpu Attempt recovering a broken GPU (unresponsive config
space or MMIO) by performing an SBR. If the GPU is
broken from the beginning and hence correct config
space wasn't saved then reenumarate it in the OS by
sysfs remove/rescan to restore BARs etc.
--reset-with-sbr Reset the GPU with SBR and restore its config space
settings, before any other actions
--reset-with-flr Reset the GPU with FLR and restore its config space
settings, before any other actions
--reset-with-os Reset with OS through /sys/.../reset
--remove-from-os Remove from OS through /sys/.../remove
--unbind-gpu Unbind GPU
--unbind-gpus Unbind GPUs
--bind-gpu=BIND_GPU Bind GPUs to the specified driver
--bind-gpus=BIND_GPUS
Bind GPUs to the specified driver
--query-ecc-state Query the ECC state of the GPU
--query-cc-mode Query the current Confidential Computing (CC) mode of
the GPU.
--query-cc-settings Query the Confidential Computing (CC) settings of the
GPU.This prints the lower level setting knobs that
will take effect upon GPU reset.
--query-prc-knobs Query all the Product Reconfiguration (PRC) knobs.
--set-cc-mode=SET_CC_MODE
Configure Confidentail Computing (CC) mode. The
choices are off (disabled), on (enabled) or devtools
(enabled in DevTools mode).The GPU needs to be reset
to make the selected mode active. See --reset-after-
cc-mode-switch for one way of doing it.
--reset-after-cc-mode-switch
Reset the GPU after switching CC mode such that it is
activated immediately.
--test-cc-mode-switch
Test switching CC modes.
--query-l4-serial-number
Query the L4 certificate serial number without the
MSB. The MSB could be either 0x41 or 0x40 based on the
RoT returning the certificate chain.
--query-module-name Query the module name (aka physical ID and module ID).
Supported only on H100 SXM and NVSwitch_gen3
--clear-memory Clear the contents of the GPU memory. Supported on
Pascal+ GPUs. Assumes the GPU has been reset with SBR
prior to this operation and can be comined with
--reset-with-sbr if not.
--debug-dump Dump various state from the device for debug
--nvlink-debug-dump Dump NVLINK debug state.
--force-ecc-on-after-reset
Force ECC to be enabled after a subsequent GPU reset
--test-ecc-toggle Test toggling ECC mode.
--query-mig-mode Query whether MIG mode is enabled.
--force-mig-off-after-reset
Force MIG mode to be disabled after a subsequent GPU
reset
--test-mig-toggle Test toggling MIG mode.
--block-nvlink=BLOCK_NVLINK
Block the specified NVLink. Can be specified multiple
times to block more NVLinks. NVLinks will be blocked
until an SBR. Supported on A100 only.
--block-all-nvlinks Block all NVLinks. NVLinks will be blocked until a
subsequent SBR. Supported on A100 only.
--dma-test Check that GPUs are able to perform DMA to all/most of
available system memory.
--test-pcie-p2p Check that all GPUs are able to perform DMA to each
other.
--read-sysmem-pa=READ_SYSMEM_PA
Use GPU's DMA to read 32-bits from the specified
sysmem physical address
--write-sysmem-pa=WRITE_SYSMEM_PA
Use GPU's DMA to write specified 32-bits to the
specified sysmem physical address
--read-config-space=READ_CONFIG_SPACE
Read 32-bits from device's config space at specified
offset
--write-config-space=WRITE_CONFIG_SPACE
Write 32-bit to device's config space at specified
offset
--read-bar0=READ_BAR0
Read 32-bits from GPU BAR0 at specified offset
--write-bar0=WRITE_BAR0
Write 32-bit to GPU BAR0 at specified offset
--read-bar1=READ_BAR1
Read 32-bits from GPU BAR1 at specified offset
--write-bar1=WRITE_BAR1
Write 32-bit to GPU BAR1 at specified offset
--ignore-nvidia-driver
Do not treat nvidia driver apearing to be loaded as an
error