GPU

Hardware → GPU shows reported GPUs on the selected node.

Read a Card

Cards show model, driver, utilization, memory, temperature, power, serial/UUID, clocks, PCIe, and interconnect readings. NVIDIA uses DCGM; AMD uses ROCm. Vendor-specific fields such as NVLink depend on hardware/driver support.

Missing values are unavailable. Empty inventory can reflect no supported GPU, exporter failure, or driver/access problems; distinguish read errors before diagnosing hardware.

ROCm 1.5.2 compatibility: implausible clocks or constant 100% utilization can result from metric mapping; compare raw gpu_clock and gpu_gfx_activity before diagnosing hardware. AMD partition changes can leave inventory stale or logical GPU identity incorrect; verify device mappings and missing readings before relying on counts or zeros.

History and Export

Export CSV includes inventory, ECC mode/counters, and throttle reasons. Disabled ECC leaves counters blank; counts depend on vendor/mode. Throttle none is an explicit empty report; a blank cell is unknown. Idle clock reduction alone is not a throttle fault.

Performance shows utilization/memory trends; Power & Thermal shows environmental readings. MAC observes installed drivers without installing them.