Gpu Hot
🔥 Real-time NVIDIA GPU dashboard
Last verified:
What is Gpu Hot?
GPU Hot is a lightweight, self-hosted real-time NVIDIA GPU monitoring dashboard that displays live GPU and system metrics directly in your web browser without requiring SSH access. Built with FastAPI and NVML on the backend and vanilla JavaScript with Chart.js on the frontend, it polls NVIDIA GPUs at 0.5-second intervals and streams data via WebSocket for sub-second metric updates. The dashboard monitors utilization, temperature, memory (used/free/total), power draw/limits, fan speed, clock speeds, PCIe info, P-State, throttle status, encoder/decoder sessions, driver/VBIOS information, and active GPU processes with PIDs and memory usage, plus host CPU, RAM, swap, disk I/O, and network throughput.
Key features include automatic multi-GPU detection that identifies every GPU in your system from consumer RTX cards to data-center A100s, historical time-series charts for utilization/temperature/power/clocks with min/max/avg stats, process-level monitoring showing which training jobs or rendering tasks consume resources, and a dual-mode hub architecture that scales from a single machine to 100+ GPUs across a cluster without extra infrastructure. The entire stack runs in a single Docker container under 200MB with one command, requires no database or external dependencies, includes nvidia-smi fallback mode for legacy Kepler-era GPUs, and offers environment-based configuration for DevOps workflows.
GPU Hot is designed for AI/ML engineers training models across multiple GPUs, DevOps professionals managing GPU infrastructure, cryptocurrency mining farm operators, university research cluster administrators, VFX/render farm coordinators, and anyone tired of constantly SSH-ing into servers to check nvidia-smi. The open-source MIT-licensed tool serves teams sharing GPU infrastructure who need to see which user or process is monopolizing resources, cloud GPU users optimizing costs by identifying underutilized instances, and homelab enthusiasts running Proxmox or Docker containers with GPU passthrough.
Gpu Hot pricing
Pricing model: Freemium
GPU Hot is completely free and open source under the MIT License. There are no paid tiers subscriptions or enterprise plans. The software is self-hosted - you deploy it yourself using Docker with the command docker run -d --gpus all -p 1312:1312 ghcr.io/psalias2006/gpu-hot:latest. All features including single-node monitoring multi-node hub mode process monitoring historical charts and the API are included at no cost. The only costs are your own infrastructure (server hardware Docker hosting electricity).
Gpu Hot pros
- Sub-second 0.5-second metric polling interval faster than nvidia-smi
- No SSH access required - browser-based monitoring
- Single Docker container deployment under 200MB
- Automatic multi-GPU detection for all NVIDIA GPUs
- Process monitoring with PID process name and memory usage
- Historical time-series charts with min/max/avg statistics
- Hub mode scales to 100+ GPUs across multiple nodes
- WebSocket real-time data streaming with Chart.js visualization
- Host system metrics including CPU RAM swap disk and network
- nvidia-smi fallback mode for legacy Kepler GPUs
- MIT license - free and open source
- Environment-based configuration for DevOps workflows
- Clean responsive UI works on any device
- No database or external dependencies required
- Per-GPU utilization temperature memory power and fan metrics
- Encoder/decoder session monitoring for video workloads
- P-State and throttle status for thermal issue detection
- PCIe info reveals slot configuration issues
- Driver and VBIOS version display
- GPU alerts at 75°C temperature or 80% utilization thresholds
Gpu Hot cons
- NVIDIA GPUs only - no AMD or Intel GPU support
- Requires NVIDIA Container Toolkit installed
- --pid=host flag needed for process names has security implications
- Default 0.5-second interval may strain hub with 50+ GPUs
- No built-in Prometheus metrics export
- No native alerting system - requires custom watchdog script
- Not suitable for production without Nginx reverse proxy authentication
- Single-port deployment may conflict if port 1312 unavailable
- No mobile app - browser access only
- Limited customization without code modifications
- No user authentication built into the dashboard
- No built-in metric retention beyond browser rolling buffer
- Documentation could be more comprehensive
- No Windows native support - Linux/Docker focused
- Hub requires manual NODE_URLS configuration
Frequently asked questions about Gpu Hot
Does GPU Hot support AMD GPUs?
No, GPU Hot is specifically designed for NVIDIA GPUs only. It uses NVIDIA's Management Library (NVML) and nvidia-smi which are NVIDIA-exclusive technologies. For AMD GPU monitoring, you should use alternatives like RadeonTop, LACT, or ROCm tools.
What is the minimum GPU architecture required?
GPU Hot works with NVIDIA GPUs dating back to Kepler architecture from 2012. For pre-Kepler cards or older GPUs without full NVML support, you must enable compatibility mode by setting the environment variable NVIDIA_SMI=true when running the Docker container.
How many GPUs can hub mode handle?
The hub has been tested with 100+ GPUs across 20 nodes. Performance depends on the hub's CPU and network latency. For larger deployments exceeding 50 GPUs, you should increase the UPDATE_INTERVAL in core/config.py from 0.5 to 1.0 seconds to reduce hub CPU usage by approximately 60%.
Is using --pid=host secure?
No, the --pid=host flag grants the container access to host process information which is a security risk. Use it only in trusted environments like personal homelabs. For production environments, consider running without it and mapping process names externally via the /api/gpu-data endpoint.
Can I export metrics to Grafana or Prometheus?
Yes, GPU Hot doesn't natively export Prometheus metrics but you can use the /api/gpu-data HTTP endpoint with a custom Prometheus exporter or Telegraf plugin. The JSON structure is stable and well-documented, returning per-GPU metrics and system stats that can be scraped and transformed.
Why aren't my metrics appearing?
First run nvidia-smi on the host to verify GPU detection. If it works, test Docker GPU access with docker run --rm --gpus all nvidia/cuda:12.1.0-base-ubuntu22.04 nvidia-smi. If that fails, install the NVIDIA Container Toolkit. For older GPUs, add the -e NVIDIA_SMI=true flag to force nvidia-smi parsing mode.
How do I update GPU Hot to the latest version?
Run docker pull ghcr.io/psalias2006/gpu-hot:latest to download the latest image, then restart your container with docker restart gpu-hot. For custom builds from source, navigate to the gpu-hot directory, run git pull to fetch updates, then docker-compose up --build to rebuild and restart.
What's the difference between worker mode and hub mode?
Worker mode runs on individual GPU servers and exposes metrics via /api/gpu-data endpoint. Hub mode aggregates data from multiple workers - you set GPU_HOT_MODE=hub and provide NODE_URLS as a comma-separated list of worker endpoints. The hub can run on a CPU-only machine without GPUs.
How do I set up custom alerts for temperature or utilization?
GPU Hot doesn't have built-in alerting, but you can create a WebSocket watchdog script that connects to ws://gpu-hot:1312/socket.io/ and sends Slack/webhook notifications when gpu.temperature exceeds 85°C or gpu.utilization exceeds 80%. The example in the documentation shows how to implement this.
Can I monitor GPU Hot on a remote server from my laptop?
Yes, deploy GPU Hot on your remote GPU server with the Docker command mapping port 1312, then access it from your laptop's browser at http://server-ip:1312. For multi-server setups, deploy workers on each GPU server and run a hub on your laptop pointing to all worker NODE_URLS for a single dashboard view.