Breaking changes
On Linux and Windows, --collect.backend now defaults to nvml instead of exec. The exporter reads the NVIDIA driver library (libnvidia-ml.so.1 on Linux, nvml.dll on Windows) directly instead of running nvidia-smi.
This change has four effects when you upgrade from 1.x:
- The collection status metric changes from
nvidia_smi_command_exit_codetonvidia_smi_nvml_return_code. - If
libnvidia-ml.so.1ornvml.dllis missing, the exporter exits at startup instead of servingnvidia_smi_last_collect_success 0. If you run the Helm chart on a mixed Kubernetes cluster, setnodeSelectorto schedule pods only on GPU nodes. - If you pass a custom
--nvidia-smi-command(such assshorsudo) or an explicit--query-field-nameslist with fields outside the NVML catalog, you must also pass--collect.backend=exec. In the Helm chart, add--collect.backend=exectoextraArgs. - If you want to keep using
nvidia-smi, pass--collect.backend=exec.
The -nvml and latest-nvml container image tags and the nvidia_gpu_exporter-nvml_* Linux archives are removed. Use the standard tags (2.0.0, latest) and the standard Linux archives (nvidia_gpu_exporter_2.0.0_linux_*.tar.gz). If you set image.tag to a -nvml tag in Helm, clear image.tag or change it to 2.0.0.
The Linux deb, rpm, and main .tar.gz artifacts now require glibc 2.34 or newer (RHEL 9, Ubuntu 22.04+, Debian 12+). If your host runs musl (such as Alpine) or an older glibc (such as RHEL 8, Ubuntu 20.04, or Amazon Linux 2), download the nvidia_gpu_exporter-static_2.0.0_linux_*.tar.gz archive instead. That static build defaults to exec.
Highlights
- Added new NVML metrics:
module.power.*fields, PCIe AER error counters (pcie_errors_total,pcie_l0_to_recovery_total), performance-policy throttle counters, signed Ada and Blackwell T.Limit thresholds, NVLink link counts and NVLink 5 recovery counters (with HGX NVSwitch fallback), C2C link counts and GPM throughput, Hopper and Blackwell GPM precision pipe activity (integer,fp64,fp32,fp16), fan control policy, repair pending, unrepairable memory, virtualization mode, and PowerMizer mode. - Added Linux PCIe runtime power management residency counters (
pcie_runtime_suspended_seconds_totalandpcie_runtime_active_seconds_total), read from sysfs without waking suspended GPUs. - Fixed Blackwell
platform.gpu_fabric_guidbyte order and omitted0xFFunavailable values onplatform.slot_numberandplatform.tray_index. - Removed the false missing-function warning on drivers older than R620.
- Added GPU captures for
NVIDIA B200,GB10,H100,A100 SXM4, andRTX 5090.
Changelog
New Features
- e1f17c9 feat!: make NVML the default backend and unify container image
- 9c9e880 feat: expand NVML and Linux runtime PM metrics
Bug Fixes
- 0bbdfd4 fix(nvml): fix platform fields, per-link recovery, and missing-fn warn
Refactors & Internal Changes
- e43c195 refactor: reduce per-scrape allocations in NVML, HTTP, and CSV paths
Dependency Updates
- 54e898b chore(deps): update actions/setup-go action to v7.1.0
- d9ed9ed chore(deps): update golang:1.27-alpine docker digest to 70bd7dd
- a6fa9aa chore(deps): update module golang.org/x/net to v0.63.0
Build, CI & Chores
- bdecd64d8e07c8a43a71ee9b92485e7dfc2d78a chore: add GB10, H100, A100 SXM4, and RTX 5090 captures
- fc58a64 chore: add NVIDIA B200 capture