NPU-Reports is an open-source repository dedicated to benchmarking, analyzing, and documenting Neural Processing Units (NPU). This project provides tools for hardware data extraction and performance reports for AI models (LLMs) on modern architectures like Intel Core Ultra, Apple Silicon, and Qualcomm Snapdragon.
- 📊 LLM Benchmarking: Real-world inference tests on NPU (Phi-3, Llama 3.1, Gemma).
- 💻 NPU-Pro Tool: A Go-based utility for hardware monitoring and control.
- 🐧 Linux Native: Optimized for Ubuntu with specific support for IPEX-LLM and Intel drivers.
- 🛡️ Cyber Analysis: Insights into hardware behavior during intensive AI workloads.
Ensure you have Go installed, then compile the main utility:
go build -o npu-pro main.go
./npu-pro --check
## 📈 Visual Performance Comparison
```mermaid
barChart
title "Inference Speed on Intel NPU (Tokens/sec)"
x-axis "Model"
y-axis "Tokens per Second"
"Llama 3.1 (8B)" : 4.17
"TinyLlama (1.1B)" : 22.0
## 📈 Visual Performance Comparison
```mermaid
pie title Inference Speed on Intel NPU (Tokens/sec)
"Llama 3.1 (8B)" : 4.17
"TinyLlama (1.1B)" : 22.0
## 🔄 OpenVINO Support (Advanced)
OpenVINO is the primary toolkit for high-performance inference on Intel hardware.
To test NPU detection via OpenVINO:
```bash
sudo apt install python3-openvino
python3 scripts/openvino/check_npu_openvino.py