nvidia-smi-webserver

A GPU metrics server that exposes temperature, power, utilization, and memory data over HTTP as JSON. Supports NVIDIA, AMD, Intel, and Apple Silicon GPUs (including the Apple Neural Engine).

Server

Requirements

• Python 3.8+
• uvicorn, fastapi, xmltodict

Running

python3 server.py

The server listens on 0.0.0.0:8000 and exposes a single endpoint:

• GET /gpu — returns an array of GPU metrics objects

API Response

[
  {
    "vendor": "nvidia",
    "index": 0,
    "name": "NVIDIA GeForce RTX 4090",
    "temperature_c": 72.0,
    "utilization_gpu_pct": 100.0,
    "memory_used_mb": 20480.0,
    "memory_total_mb": 24564.0,
    "power_draw_w": 350.0
  }
]

---

macOS / Apple Silicon

On Apple Silicon (M1/M2/M3/M4 and variants) the server reports two entries with "vendor": "apple":

• Integrated GPU (index: 0) — name, GPU busy %, power draw, and unified memory usage.
• Neural Engine / ANE (index: 1) — power draw.

Metric sources & caveats

• GPU name and core count come from system_profiler and work without root.
• Live metrics (GPU busy %, GPU power, ANE power) come from powermetrics, which requires root. Run the server with sudo to populate these fields; otherwise the device is still listed by name with dynamic fields set to null.
• Unified memory: Apple Silicon shares one memory pool between CPU, GPU, and ANE, so there is no dedicated VRAM. memory_total_mb is the total unified memory and memory_used_mb is the system-wide used memory (active + wired + compressed), since that is the pool the GPU draws from.
• Temperature is not exposed through a root-free API on Apple Silicon, so temperature_c is null.

# For full metrics on macOS, run as root:
sudo python3 server.py

---

Deploy the server as a systemd service on the GPU machine so it starts on boot and auto-restarts on failure.

Setup

# Copy files to the GPU machine
sudo mkdir -p /opt/gpu-monitor
sudo cp server.py /opt/gpu-monitor/
sudo cp deploy/gpu-monitor.service /etc/systemd/system/

# Install Python dependencies
sudo pip3 install uvicorn fastapi xmltodict

# Enable and start
sudo systemctl daemon-reload
sudo systemctl enable gpu-monitor.service
sudo systemctl start gpu-monitor.service

# Check status
sudo systemctl status gpu-monitor.service
journalctl -u gpu-monitor.service -f

Firewall

Ensure port 8000 is accessible from your development machine:

sudo ufw allow 8000/tcp

---

VSCode Extension — GPU Thermal Throttle Monitor

A VSCode extension that monitors a remote GPU server and alerts you when thermal throttling is detected.

How It Works

1. Polls the GPU server at a configurable interval
2. Tracks temperature, power draw, and GPU utilization
3. Uses a state machine to detect throttling:
   • NORMAL → HOT: Temperature hits threshold while power is high (GPU under full load at thermal limit)
   • HOT → THROTTLING: Power draw drops while temperature stays high and utilization is still 100% (GPU is clock-throttling to shed heat)
   • Recovery: Temperature drops below threshold minus 3°C hysteresis

Notifications

• Warning (HOT): "GPU temperature has reached 87°C at 350W — near thermal limit"
• Error (THROTTLING): "GPU appears to be thermally throttling — power dropped from 350W to 280W while utilization remains at 100%. Performance is degraded."

Notifications have a configurable cooldown (default: 60s) to avoid spam.

Status Bar

Shows real-time GPU stats in the VS Code status bar:

• 🟢 ✓ GPU: 65°C | 320W | 98% — Normal
• 🟡 ⚠ GPU: 87°C | 350W | 100% — Hot
• 🔴 🔥 GPU: 89°C | 280W | 100% — Throttling

Installation

cd vscode-extension
npm install
npm run compile

Then press F5 in VS Code to launch the extension in a development host, or package it:

npx vsce package
theia --install-extension gpu-thermal-monitor-0.1.0.vsix

Configuration

Open VS Code Settings and search for "GPU Monitor":

Setting	Default	Description
gpuMonitor.serverUrl	http://localhost:8000	URL of the GPU monitoring server
gpuMonitor.pollIntervalMs	2000	Polling interval in ms
gpuMonitor.temperatureThreshold	87	Temperature (°C) to trigger alerts
gpuMonitor.powerHighThreshold	300	Power (W) considered "high" at full load
gpuMonitor.powerDropPercent	15	% power drop that indicates throttling
gpuMonitor.gpuIndex	0	Which GPU to monitor
gpuMonitor.notificationCooldownMs	60000	Minimum time between repeated alerts

Commands

• GPU Monitor: Start Monitoring — begin polling
• GPU Monitor: Stop Monitoring — stop polling
• GPU Monitor: Show GPU Status — display current state summary

---

License

See LICENSE.