GPU

Per-GPU live state for the picked node. The page is a flat grid of cards — one per GPU — showing every metric the dashboard's GPU drilldowns expose.

Per-Card Fields

Field	What it means
Name	Marketing name (e.g. RTX PRO 6000 Blackwell Server Edition)
Driver	Vendor driver version reported by DCGM / ROCm
Utilization	SM / shader-engine activity (0–100 %)
Memory	Used / total VRAM, with a usage bar
Temperature	Hottest die sensor (°C); colored when over threshold
Power	Current draw (W), with the configured cap on the right
Serial	Vendor-reported GPU serial number
UUID	Stable hardware UUID (GPU-...)
Graphics Clock / Memory Clock	Live clock rates
Throttle	DCGM throttle reasons; None when the GPU is unconstrained
ECC Errors	Single-bit and double-bit error counters (NVIDIA)
PCIe Link	Negotiated speed × width
NVLink BW	Live NVLink bandwidth (when applicable)

Export CSV

The button in the upper-right exports the current cards into a CSV — handy for spreadsheets, inventory audits, and procurement.

Multi-Vendor Notes

• NVIDIA GPUs source data from dcgm-exporter; almost every field is filled.
• AMD GPUs source data from rocm-exporter; fields like NVLink BW and Throttle are vendor-specific and shown as —.

See Also

• Performance — time-series view of utilization, memory, power
• Power & Thermal — heatmap and rolling thermal trace