Linux NVIDIA GPUs: Match the Physical Slot and PCI Bus Address
1. Inspect SMBIOS slot information
The source uses this command to inspect slot designation and PCI bus address:
dmidecode -t slot
This example shows SLOT1, ID 1, in use, and 0000:3b:00.0. SMBIOS data quality and slot mapping depend on firmware/platform; correlate the actual server rather than assuming a universal bus-to-slot map.
2. Inspect NVIDIA PCI devices
lspci | grep -i nvidia
The example lists NVIDIA device 1eb8, revision a1, at 3b:00.0, 3c:00.0, 5e:00.0, 86:00.0, 87:00.0, af:00.0, d8:00.0, d9:00.0. The first matches the shown slot record. A numeric PCI device ID alone may not supply the complete marketed GPU name.
3. Inspect the driver-reported GPU name and bus ID
nvidia-smi
This image is an English explanatory reconstruction of the source’s annotated example, not a newly collected terminal screenshot. Its June 26, 2017 output has driver/NVIDIA-SMI 384.37, two Tesla M60 rows and a truncated Tesla P100-PCIE name, at buses 0000:04:00.0, 0000:05:00.0 and 0000:06:00.0. Those are a different example from the eight-device lspci list and must not be treated as the same system.
The diagram explains index, GPU name, persistence mode, bus ID, display/ECC state, temperature, performance state, power usage/limit, used/total framebuffer memory, utilization and compute mode. Its process example is GPU 2, PID 1933, C+G and /usr/lib64/xen/bin/vgpu. The process-memory digits are cropped in the source and remain 81...MiB; the original annotation says an 8 GB vGPU allocation. No missing digits are reconstructed.
For the real machine, match the full PCI domain:bus:device.function across commands. A GPU index is not a permanent physical identifier. Driver-reported state and accurate SMBIOS/platform documentation are both needed to select a physical service location.
English references: NVIDIA System Management Interface documentation.