Tink vs NVIDIA GPU Exporter
nvidia_gpu_exporter is Prometheus's usual nvidia-smi scrape: a :9835 /metrics endpoint of utilization, VRAM, and temperature gauges. That is not a mechanic. Tink installs in one command, watches the Linux server, explains what is wrong, and helps you fix it.
nvidia_gpu_exporter exposes GPU gauges. Tink answers “why is this VPS sick?”
Accidental sysadmins land on nvidia_gpu_exporter because every GPU + Prometheus diagram starts with an nvidia-smi scrape. Then the volume fills, nginx dies, and Grafana is quiet — the card still sits at 12%, so nothing fired. A filling /var is a Tink warning; it is not an exporter job.
Tink is the other job: detect the issue on the machine, say why it happened in plain English, propose the command, and run it only after you approve. Keep nvidia_gpu_exporter if you already operate Prometheus and need GPU gauges. Use Tink if you run Linux servers and want a mechanic.
| Feature | Tink | NVIDIA GPU Exporter |
|---|---|---|
| Setup time | 30 seconds (one curl | sh command) | Hours — NVIDIA drivers, nvidia-smi, :9835, prometheus.yml scrape job, dashboards, and Alertmanager |
| What you get | Working monitoring, diagnosis, and approved fixes | A GPU scrape target. Host disk fill, nginx, Postgres, and certs stay your problem until PromQL fires and you SSH in |
| Pricing | Free (Scout) / $9 / $29 per machine per month | Free software (MIT). You still pay for Prometheus, Grafana, Alertmanager, and the engineer who writes the rules |
| Hidden costs | None — fully managed | Driver versions, MIG vs full-card flags, a noisy temperature series, and a second tool to tell you the box is sick |
| Monitoring approach | Agent on the server — CPU, disk, services, logs, certs, ports | nvidia_smi_utilization_gpu and VRAM gauges. nvidia_gpu_exporter does not watch host fill or a quiet cert unless you added another exporter |
| Configuration | None after install — heuristics and AI | --query-field lists, GPU index filters, scrape intervals, and which PromQL actually pages you |
| Plain-English diagnosis | Yes — AI explains root cause, impact, and fix | You graph nvidia_smi_temperature_gpu in Grafana, then SSH in to change the box |
| Fix execution | Proposes and executes approved commands with an audit trail | Export only — nvidia_gpu_exporter cannot free disk, restart nginx, or rotate a cert |
| Alerting | Built-in across 8 channels | Not an NVIDIA GPU Exporter job — Prometheus rules and Alertmanager have to fire |
| Predictive alerts | Yes — disk fills in ~6 days, memory and CPU trends | Not an exporter job — a GPU scrape does not forecast a quiet volume filling up |
| SSH brute-force detection | Built-in — parses auth.log every scan | Only if you ship auth.log elsewhere, parse it, write a query, and wire an alert downstream |
| Machine offline detection | Agent presence monitoring with multi-channel alerts | Silence if nvidia_gpu_exporter dies — unless you built an up{job="nvidia-gpu"} deadman check |
| Public status page | Shareable URL with 90-day history | None — nvidia_gpu_exporter has no customer-facing status page |
| Weekly fleet digest | Automated Monday digest + daily brief when issues are open | None — a GPU scrape target is not a plain-English fleet narrative |
| On-call tracking | Built-in /oncall command + incident acknowledgment | Not included — wire Alertmanager into PagerDuty or another incident tool |
| Conversation interface | Telegram, WhatsApp, web dashboard, CLI | curl :9835/metrics and Grafana. No mechanic you text when /var filled and the GPU still said 12% |
| Learning curve | None — works after install | Steep — nvidia-smi, query fields, PromQL, scrape configs, and which rule actually pages |
| Best for | Freelancers, small teams, accidental sysadmins (1-50 Linux servers) | Teams that already run Prometheus and need GPU scrapes, not a sick-VPS mechanic |
When NVIDIA GPU Exporter is the right choice
Keep nvidia_gpu_exporter when you already want that exact job:
- Prometheus GPU scrapes — you need nvidia_smi_utilization_gpu in PromQL, not a mechanic.
- Per-card hardware gauges — VRAM, power draw, and GPU temperature matter more than host nginx, certs, and SSH on a quiet VPS.
- An existing Prometheus + Grafana stack — nvidia_gpu_exporter already ships, dashboards already graph, and someone owns Alertmanager.
- Hosts that only export — you already run nvidia_gpu_exporter next to the cards, not a sick-box SSH loop.
The real cost of “just run nvidia_gpu_exporter”
A GPU exporter is in every nvidia-smi + Prometheus diagram. A working ops loop for five Linux boxes is not:
- An exported 12% GPU still means you SSH in and change the box by hand
- Host fill, nginx, certs, and SSH brute-force never appear unless you scraped those signals and Alertmanager paged
- A filling /var is a Tink disk warning on every scan — nvidia_gpu_exporter will keep scraping idle VRAM
- A single VPS with a dead nvidia_gpu_exporter still needs a mechanic, not another scrape job
For a 5-server team, Tink Mechanic at $45/month is cheaper than nvidia_gpu_exporter plus the engineer who keeps Prometheus scrape alive, and every scan includes a diagnosis a quiet /metrics endpoint will not type.
No prometheus.yml homework. No silent host. One command install.
Start MechanicAlso compare: Tink vs Node Exporter · Tink vs Prometheus · Tink vs Process Exporter