UOMI Network GPU Benchmark

A tool for evaluating GPU hardware suitability for AI inference in the UOMI Network.

Overview

This benchmark evaluates NVIDIA GPUs to determine if they meet the requirements for participating in the UOMI Network's distributed AI inference system. It performs comprehensive testing of:

• Computational performance (TFLOPS)
• Memory capacity and bandwidth
• Multi-GPU scaling efficiency
• Overall suitability for LLM inference

Quick Start

Run the benchmark with a single command:

curl -sSL https://raw.githubusercontent.com/Uomi-network/uomi-gpu-benchmark/refs/heads/main/benchmark.sh | bash

The benchmark will:

1. Check system requirements
2. Install necessary Python dependencies (in a virtual environment if needed)
3. Run comprehensive GPU performance tests
4. Generate a detailed report with your hardware score and classification

Requirements

• NVIDIA GPU with CUDA support (minimum 16GB VRAM recommended)
• Ubuntu/Debian Linux (other distros should work but are less tested)
• Python 3.6+ with pip
• Internet connection to download dependencies

For optimal performance in the UOMI Network, we recommend:

• 2 x NVIDIA RTX 4090 or better ( score: 800)
• 48GB+ VRAM

Scoring System

The benchmark uses a 1000-point scoring system to evaluate hardware:

Tier	Score Range	Classification	Examples
S+	2000+	Data Center	H100 SXM5, Multiple A100s
S	1100-1999	Professional	H100 PCIe, Multiple RTX 4090s
A+	900-1099	High-End	2x RTX 4090, H100 PCIe
A	750-899	Premium	A100 80GB, 2x RTX 3090
B+	550-749	Performance	A100 40GB, RTX 4090
B	450-549	Mainstream	A6000, RTX 3090
C	300-449	Entry	A4000, RTX 3080
D	0-299	Minimal	Consumer GPUs <16GB

LLM Inference Capabilities

The benchmark also estimates large language model (LLM) capabilities:

Classification	Parameter Range	Example Models
SUPERB	180B+	BLOOM 176B, GPT-3 175B
EXCELLENT	100-180B	Llama 2 70B (high throughput)
GREAT	70-100B	Claude Opus, GPT-4, Llama 2 70B
VERY GOOD	40-70B	Falcon 40B, Llama 2 70B (quantized)
GOOD	20-40B	Llama 2 30B, Claude Sonnet
FAIR	10-20B	Llama 2 13B, MPT 30B (quantized)
BASIC	<10B	Mistral 7B, Phi-2

Interpreting Results

After running the benchmark, you'll see output similar to:

📊 GPU SCORE: 877/1000
🏆 TIER: B+ (Performance)
💾 Total GPU Memory: 47.29 GB
⚡ Combined Performance: 102.01 TFLOPS
🧠 Estimated LLM Parameters: up to 31B

🏁 FINAL RESULT: 877/1000 - TIER B+ (Performance)
✅ HARDWARE ACCEPTED for UOMI network

A detailed JSON file gpu_benchmark_result.json will also be saved with comprehensive metrics.

Multi-GPU Support

The benchmark automatically detects and tests all available NVIDIA GPUs. It provides:

• Individual performance metrics for each GPU
• Combined performance estimates with appropriate scaling factors
• Optimized scoring for both homogeneous (same model) and heterogeneous (mixed) setups

Troubleshooting

Dependency Issues

If you encounter Python package installation problems:

# On Ubuntu/Debian
sudo apt install python3-venv python3-pip

Permission Errors

If you get permission errors:

# Run with sudo privileges
sudo curl -sSL https://raw.githubusercontent.com/uomi-network/gpu-benchmark/main/benchmark.sh | sudo bash

CUDA Not Found

If the benchmark can't detect your GPU:

1. Ensure you have the NVIDIA drivers installed
2. Check that CUDA is installed and in your PATH
3. Verify the GPU is properly seated and powered

Contributing

Contributions are welcome! Please feel free to submit a pull request.