gpu-benchmark-tool

A comprehensive multi-vendor GPU health monitoring and optimization tool that helps users assess GPU performance and select optimal hardware for their workloads.

🚀 Features

🔥 Comprehensive GPU Health Monitoring: Temperature, power, utilization, and throttling detection

⚡ Advanced Stress Testing: Compute, memory bandwidth, VRAM, and mixed-precision tests

📊 Detailed Health Scoring: 100-point scoring system with actionable recommendations

🖥️ Multi-GPU Support: Test and compare multiple GPUs simultaneously

🧪 Mock Mode: Test on any computer without GPUs (perfect for development)

🔌 Multi-Vendor Support: NVIDIA, AMD, Intel, and Mock mode

☁️ Cloud-Ready: Designed to help select optimal GPUs for cloud deployment (coming soon!)

Installation

Basic Installation (Works on any system with GPU)

For systems with any GPU (NVIDIA, AMD, Intel)

pip install gpu-benchmark-tool

Includes PyTorch for computational stress tests

Installation with Enhanced GPU Support

For NVIDIA GPUs (adds NVIDIA monitoring)

pip install gpu-benchmark-tool[nvidia]

For AMD GPUs (relies on system ROCm)

pip install gpu-benchmark-tool[amd]

For Intel GPUs (adds Intel GPU acceleration)

pip install gpu-benchmark-tool[intel]

For all GPU vendors (maximum compatibility)

pip install gpu-benchmark-tool[all]

🎯 Quick Start

• Check Available GPUs gpu-benchmark list

• Run Benchmark

Benchmark all GPUs

gpu-benchmark benchmark

Benchmark specific GPU (recommended)

gpu-benchmark benchmark --gpu-id 0

Quick 30-second test

gpu-benchmark benchmark --gpu-id 0 --duration 30

Export results to JSON

gpu-benchmark benchmark --gpu-id 0 --export results.json

• Mock Mode (No GPU Required)

Perfect for development or systems without GPUs

gpu-benchmark benchmark --mock --duration 30

📊 Google Colab Quick Start

Run in a Colab notebook (Runtime > Change runtime type > GPU)

!pip install gpu-benchmark-tool[nvidia] !gpu-benchmark benchmark --gpu-id 0 --duration 30

Understanding Results

Health Score (0-100 points) 85-100: 🟢 Healthy - Safe for all workloads including AI training 70-84: 🟢 Good - Suitable for most workloads 55-69: 🟡 Degraded - Limit to inference or light compute 40-54: 🟡 Warning - Monitor closely, avoid heavy workloads 0-39: 🔴 Critical - Do not use for production

Score Components

Each component contributes to the total 100-point score:

Temperature (20 points)

• Peak temperature during stress test
• Under 80°C: Full points
• 80-85°C: 15 points
• 85-90°C: 10 points
• Over 90°C: 5 points

Baseline Temperature (10 points)

• GPU temperature at idle
• Under 50°C: Full points
• 50-60°C: 5 points
• Over 60°C: 0 points

Power Efficiency (10 points)

• Power consumption optimization
• Within optimal range: Full points
• Slightly outside range: 5 points
• Far from optimal: 0 points

GPU Utilization (10 points)

• How well the GPU is utilized during tests
• 99%+: Full points
• 90-98%: 5 points
• Under 90%: 0 points

Throttling (20 points)

• Thermal or power throttling detection
• No throttling: Full points
• Occasional throttling: 10-15 points
• Frequent throttling: 0-5 points

Errors (20 points)

• Stability during stress tests
• No errors: Full points
• Few errors: 10-15 points
• Many errors: 0-5 points

Temperature Stability (10 points)

• Temperature consistency during tests
• Very stable: Full points
• Some fluctuation: 5-7 points
• Unstable: 0-5 points

Performance Metrics

Matrix Multiplication: Raw compute performance (TFLOPS) Memory Bandwidth: Memory throughput (GB/s) VRAM Stress: Memory allocation stability Mixed Precision: FP16/BF16 support for AI workloads

Command Line Usage

Benchmark Command

gpu-benchmark benchmark [OPTIONS]

Options: --gpu-id INTEGER Specific GPU to test (default: all GPUs) --duration INTEGER Test duration in seconds (default: 60) --basic Run basic tests only (faster) --export TEXT Export results to JSON file --verbose Show detailed output --mock Use mock GPU (no hardware required)

Examples

Full test on GPU 0 with export

gpu-benchmark benchmark --gpu-id 0 --duration 120 --export full_test.json

Quick health check

gpu-benchmark benchmark --gpu-id 0 --duration 30 --basic

Development testing

gpu-benchmark benchmark --mock --export mock_results.json

Real-time Monitoring

Monitor GPU metrics in real-time (NVIDIA only)

gpu-benchmark monitor --gpu-id 0

Python API Usage

Basic Usage

import pynvml from gpu_benchmark import run_full_benchmark

Initialize NVML

pynvml.nvmlInit() handle = pynvml.nvmlDeviceGetHandleByIndex(0)

Run benchmark

results = run_full_benchmark( handle=handle, duration=60, enhanced=True, device_id=0 )

Access results

print(f"Health Score: {results['health_score']['score']}/100") print(f"Status: {results['health_score']['status']}")

Analyzing Results

Check if GPU is healthy for production

if results['health_score']['score'] >= 70: print("✅ GPU is suitable for production workloads") else: print("⚠️ GPU needs attention")

Access performance metrics

if 'performance_tests' in results: tflops = results['performance_tests']['matrix_multiply']['tflops'] print(f"Compute Performance: {tflops:.2f} TFLOPS")

🔧 Troubleshooting

Common Issues

"No GPUs found"

Use --mock flag for testing without GPUs Ensure NVIDIA/AMD/Intel drivers are installed For AMD: Install ROCm drivers and PyTorch with ROCm support For Intel: Install Intel GPU drivers and Intel Extension for PyTorch

"NVML Error" on Colab

This warning can be ignored - the tool still works correctly Use --gpu-id 0 for cleaner output

"PyTorch not available"

The base installation now includes PyTorch If you see this error, try: pip install gpu-benchmark-tool[nvidia]

Low Health Scores

Check system cooling Ensure GPU isn't thermal throttling Close other GPU applications Multi-GPU JSON Format

Use --gpu-id 0 to test single GPU (simpler output) Without --gpu-id, results are nested under 'results' key

Supported GPUs

NVIDIA GPUs (Full Support) Consumer: RTX 4090, 4080, 4070, 3090, 3080, 3070, 3060 Data Center: A100, V100, T4, P100, K80 Workstation: RTX A6000, A5000, A4000 AMD GPUs (ROCm Required) MI250X, MI210, MI100 Radeon RX 7900 XTX, RX 6900 XT Intel GPUs (Limited Support) Arc A770, A750 Intel Xe integrated graphics

Requirements

Python 3.8 or higher For NVIDIA: CUDA drivers For AMD: ROCm drivers For Intel: Intel GPU drivers

📄 License MIT License - see LICENSE file for details.

🙏 Acknowledgments Built to solve real-world GPU selection challenges and reduce cloud computing costs through better hardware decisions.

📧 Contact PyPI: https://pypi.org/project/gpu-benchmark-tool/ Email: [e-mail]