gpu

Documentation

Overview

Package gpu samples accelerator telemetry across vendors.

NVIDIA/AMD/Intel are read through their vendor CLIs (nvidia-smi, rocm-smi, xpu-smi). We shell out deliberately: NVML and Level Zero have no stable in-process Go API without cgo-linking driver libraries, and the vendor CLIs are their documented interfaces. Tool presence is cached for the process lifetime.

Index

• func ParseNvidiaSMI(b []byte) []core.GPUDevice
• func ParseRocmSMI(b []byte) []core.GPUDevice
• func Sample(ctx context.Context) []core.GPUDevice

Functions

func ParseNvidiaSMI

func ParseNvidiaSMI(b []byte) []core.GPUDevice

ParseNvidiaSMI reads CSV rows of index,name,temp,memused,memtotal,util,power,driver_version. The name may contain commas; the last six fields are always fixed. Exported so the remote ssh path can parse nvidia-smi output gathered from another host with the exact same rules.

func ParseRocmSMI

func ParseRocmSMI(b []byte) []core.GPUDevice

ParseRocmSMI reads rocm-smi --json output keyed by "cardN". Values may be strings or single-element arrays depending on version. Exported so the remote ssh path can parse rocm-smi output gathered from another host.

func Sample

func Sample(ctx context.Context) []core.GPUDevice

Sample collects devices from every vendor present on the host.