This repository has the resources to perform tests on Copilot+ PCs, which contain NPUs (Neural Processing Unit).

The NPU (Neural Processing Unit) is a specialized processor designed to accelerate artificial intelligence and machine learning tasks.

An AI model "fits" on a device if it can run locally (without relying on the cloud) using the device's NPU. To achieve this, three conditions must be met:

1. Sufficient calculation power (TOPS).
2. Available memory (Sufficient RAM to load model weights).
3. Model optimization specific to the NPU architecture.

TOPS (Trillions of Operations Per Second) is the performance metric used to quantify an NPU's calculation capacity.

Relation between TOPS and AI Models:

• ~10 TOPS (Small Models): Sufficient for image classification, simple object detection, or basic assistants.
• >40 TOPS (Large Models): Required to fluidly run text generation models (like GPT), advanced image editing, automatic translation, or multimodal assistants.

• CPU: Qualcomm® Oryon™ — ARM64 architecture, 8 cores, max frequency 3.42 GHz, designed for high performance and power efficiency on mobile and desktop platforms with Windows on ARM.
• Integrated GPU: Qualcomm® Adreno™ X1-85 — compatible with DirectX 12 and DirectML, optimized for advanced graphics tasks and AI acceleration.
• NPU (Qualcomm Hexagon): 45 TOPS (INT8), focused on local generative AI workloads, capable of running complex language, vision, and audio models without cloud dependency.
• Max General TOPS (INT8): 45 TOPS (Dedicated NPU performance, without aggregated CPU/GPU compute).
• Memory: 32 GB LPDDR5X — maximum capacity supported by the platform, optimized for efficiency and high bandwidth.
• Storage: 954 GB.
• OS: Windows 11 Enterprise (version 24H2).

• Commercial Name: Part of the Qualcomm Hexagon NPU.
• Performance: 45 TOPS in INT8.
• Design: Centered on local generative AI, prepared to run complex language, vision, and audio models without the need for the cloud.
• Ecosystem: Supports ONNX Runtime with DirectML and the Qualcomm AI Engine SDK.

• CPU: Intel® Core™ Ultra 5 228V — 8 cores (4 P-cores + 4 E-cores), Turbo up to 4.5 GHz, base frequency P-core 2.1 GHz, 8 MB cache. Base power 17 W / turbo 37 W.
• Integrated GPU: Intel® Arc™ 130V (7 Xᵉ-cores), dynamic graphics frequency up to 1.85 GHz, GPU TOPS (INT8): 53. Supports ray tracing and AV1/H.264/H.265 codecs.
• NPU (Intel AI Boost): 40 TOPS (INT8), supports sparsity and is integrated into the platform.
• Max General TOPS (INT8): 97 TOPS (Aggregated platform value reported in spec sheet).
• Memory: 32 GB installed (CPU max support: 32 GB LPDDR5X up to 8533 MT/s).
• Storage: 477 GB.
• OS: Windows 11 Enterprise (version 24H2).

Note: For specific, step-by-step installation instructions, please refer to the README.md file located inside each device's specific folder in this repository.

Installations differ fundamentally because the processors possess distinct architectures, which directly affects the compatibility of software, frameworks, and libraries. Each architecture requires its own execution environment, preventing cross-installation.

• Snapdragon X Elite (ARM64):

  • Uses ARM64 architecture.
  • Incompatibility: It is not possible to install OpenVINO, as this framework is designed to run on x64/x86 architectures.
  • Python Requirement: Execution on the NPU requires a version of Python compiled specifically for ARM64.

• Intel Core Ultra (x64):

  • Uses x64 (x86-64) architecture.
  • Incompatibility: It is not possible to install an ARM64 version of Python. This prevents the use of libraries such as QNN (Qualcomm Neural Network) within ONNX Runtime, which are designed specifically for ARM64 processors.