AI Power Dashboard ⚡

Self-hosted NVIDIA GPU energy and Strata token monitoring for Linux AI workstations.

Track multiple NVIDIA GPUs, estimated whole-system energy usage, electricity bills, and the API-equivalent value of tokens generated locally — using a lightweight Python web dashboard. Designed for always-on LLM servers running tools such as Strata and OpenCode.

No Docker, Nginx, cloud account, JavaScript build tools, or Python packages required.

Features

• Multi-GPU power monitoring: per-GPU watts, utilization, VRAM and energy (kWh) from nvidia-smi, identified by GPU UUID.
• Estimated wall power: GPU readings + configurable CPU idle/load estimates + motherboard/RAM/storage/fans estimate, adjusted for PSU efficiency.
• Electricity costs: editable price per kWh and currency, plus historical energy totals.
• Strata token analytics: input (totals.prompt_tokens), output (totals.output_tokens) and reused/cached input (totals.reused), using cumulative counter deltas.
• API cost comparison: editable USD input/output prices per million tokens. This is a hypothetical API-equivalent cost, not a precise comparison with subscription plans or model quality.
• Historical SQLite storage: 24h, 7d, 30d and all-time reporting, surviving reboots and service restarts.
• Live web UI: auto-updates every 5 seconds, with a power chart, GPU details and settings.
• Boot-time startup: optional systemd service.

Requirements

• Ubuntu or another Linux distribution with Python 3.9+ and systemd (for the installer).
• Working proprietary NVIDIA driver and nvidia-smi, with power reporting supported by the installed GPUs.
• Optional Strata HTTP metrics endpoint returning JSON (default: http://127.0.0.1:8080/metrics).

Hardware notes: The initial CPU/board defaults are illustrative estimates for an Intel Core i7-6850K / X99 workstation: CPU idle 25 W, CPU busy 110 W, other components 55 W, PSU efficiency 88%. They are not hardware measurements. Intel's CPU TDP is not equal to actual wall power. For accurate whole-PC costs, use a calibrated wall power meter or smart plug. This project does not yet integrate smart plugs.

Quick start

git clone https://github.com/egortar-pi/ai-power-dashboard.git
cd ai-power-dashboard
nvidia-smi  # Verify the driver and all cards first
chmod +x install.sh
./install.sh

Open the dashboard from the same machine at http://127.0.0.1:8092.

🌐 Accessing the Dashboard over LAN

By default, AI Power Dashboard listens on 127.0.0.1:8092, meaning it is only accessible from the machine on which it is running.

To access the dashboard from other devices on your local network, change the listening address to 0.0.0.0.

1. Update the systemd service

Open the service configuration:

sudo nano /etc/systemd/system/ai-power-dashboard.service

In the [Service] section, set:

Environment=HOST=0.0.0.0

Note: This assumes your installation reads the HOST environment variable. If the application uses a command-line argument instead, change its host argument to --host 0.0.0.0.

2. Restart the service

sudo systemctl daemon-reload
sudo systemctl restart ai-power-dashboard

Verify that the dashboard is listening on all network interfaces:

sudo ss -lntp | grep 8092

Expected output should include:

0.0.0.0:8092

3. Configure the firewall (optional)

If UFW is enabled, allow access from your local network:

sudo ufw allow from 192.168.0.0/24 to any port 8092 proto tcp

Replace 192.168.0.0/24 with your actual LAN subnet.

4. Open the dashboard

From another device on your network, visit:

http://YOUR_SERVER_IP:8092

For example:

http://192.168.0.152:8092

🔒 Security Notice

Binding to 0.0.0.0 makes the dashboard accessible through all network interfaces, not just your LAN.

The dashboard does not currently provide built-in authentication, and its settings can be modified through the web interface.

• Use this configuration only on trusted networks.
• Restrict access using firewall rules.
• Do not expose port 8092 directly to the public internet.
• For remote access, prefer SSH tunneling, Tailscale, or a secured reverse proxy.

For maximum security, keep the default 127.0.0.1 binding and access the dashboard through an SSH tunnel.

Manual run (without systemd)

python3 app.py serve
# Other terminal: python3 app.py report --period 7d

Configuration

Open Settings in the dashboard to edit:

Strata's cached/reused input token counter is read from totals.reused. Reused tokens are already included in prompt tokens; never add them a second time. Current API comparison treats all prompt tokens equally and does not apply cached-token discounts.

Changes to the electricity rate recalculate the displayed historical energy cost using the current rate, not the rate applicable when the electricity was consumed. For time-of-use tariffs or historical tariff changes, a rate-history feature would be needed.

Data and calculations

GPU power: queried from nvidia-smi for every discovered GPU. Adjacent samples are integrated using the trapezoidal rule (W × elapsed time → kWh). Power is not reconstructed through monitoring outages; gaps beyond two sampling intervals are capped.

Estimated system power:

CPU_W = CPU_IDLE_W + CPU_utilization * (CPU_BUSY_W - CPU_IDLE_W)
WALL_W ≈ (SUM(GPU_W) + CPU_W + OTHER_W) / PSU_EFFICIENCY
ENERGY_COST = SYSTEM_KWH * ELECTRICITY_PRICE_PER_KWH

CPU utilization comes from /proc/stat, not package power telemetry. If GPU power is unavailable, the collector logs a warning and pauses whole-system energy accumulation rather than falsely treating GPU use as zero. This conservative accounting undercounts electricity during outages; review error messages and restore nvidia-smi to resume measurements.

Token counts: cumulative Strata counters are differenced at each poll. On the initial observation or a counter reset/restart, a new baseline is created. Tokens generated before monitoring started are not backfilled or invented.

GPU identification: the NVIDIA UUID is used to keep card histories consistent even when index ordering changes after reboot.

Storage: SQLite lives in ~/.local/share/ai-power-dashboard/history.sqlite3 for the service user, unless AI_DASH_DIR is configured. Protect it as potentially sensitive operational data.

Operations

sudo systemctl status ai-power-dashboard
sudo systemctl restart ai-power-dashboard
journalctl -u ai-power-dashboard -f
python3 app.py sample
python3 app.py report --period 7d
python3 -m unittest discover -s tests -v

To choose a different binding or port, edit the service's Environment=AI_DASH_HOST=127.0.0.1 / Environment=AI_DASH_PORT=8092 entries, then sudo systemctl daemon-reload && sudo systemctl restart ai-power-dashboard.

Security

No authentication is built in. Settings can be changed via the API. By default the dashboard listens only on 127.0.0.1. Use SSH port forwarding or a trusted VPN. Binding to 0.0.0.0 exposes it to the network: only do so in a fully trusted LAN, restrict ingress with a firewall, and do not forward its port to the public internet.

The dashboard does not collect prompts, responses or API keys, but may store filenames, GPU identifiers, operational timestamps and usage history. Do not commit the SQLite database or private configuration.

Troubleshooting

Failed to initialize NVML: Driver/library version mismatch means the loaded NVIDIA kernel driver does not match the userspace library. After a driver upgrade, a planned reboot often resolves this. Check nvidia-smi before troubleshooting the dashboard itself.

Strata tokens stay at zero: verify curl http://127.0.0.1:8080/metrics and inspect the totals object. The first successful poll establishes a baseline, so prior requests are not included.

API-equivalent value vs electricity mismatch: API comparisons are priced in USD; electricity is shown in your chosen currency. No exchange rate is assumed or silently applied.

Limitations / roadmap

• CPU, motherboard, RAM and PSU consumption are estimated; wall-meter integration is not implemented.
• If any GPU reading fails, the affected collection interval can be missing from cost totals.
• Dashboard reports all energy used by the monitored workstation, not energy attributable specifically to one AI job.
• The Strata collector currently reads one Strata endpoint. Other OpenCode backends require separate integrations.
• Subscription savings, break-even analyses, cached-token API pricing, multi-engine accounting and tariff history are potential future improvements.

Contributing / license

See CONTRIBUTING.md. Distributed under the MIT License.