AsNumpy is a lightweight Python library for scientific computing on Ascend NPU, fully compatible with the NumPy API. It wraps Huawei CANN operators through a pybind11 binding layer, exposing them via the NPUArray data structure that mirrors numpy.ndarray. Developed by the AISS Group and the ISE Group at Harbin Institute of Technology, in collaboration with the Huawei CANN team.

import numpy as np
import asnumpy as ap

# Generate data on CPU
m1 = np.random.normal(0, 1, (3000, 3000)).astype(np.float32)
m2 = np.random.normal(0, 1, (3000, 3000)).astype(np.float32)

# Transfer to NPU
m1_npu = ap.ndarray.from_numpy(m1)
m2_npu = ap.ndarray.from_numpy(m2)

# Compute on NPU — same API as NumPy
result = ap.multiply(m1_npu, m2_npu)

# Transfer back to CPU
print(result.to_numpy())

Note: The Performance section below benchmarks ap.mean() (a reduction operation), which differs from the ap.multiply() (element-wise) demo above. See benchmarks.md for full reproduction steps.

• Features
• Installation
• Quick Start
• Performance
• Roadmap
• Contributing
• Resources
• License
• Acknowledgements

Features

• NumPy-compatible API — function names and signatures match NumPy; migrate existing code with minimal changes
• Ascend 910B native acceleration — operators run directly on NPU via CANN ACLNN without framework overhead
• Automatic resource management — NPUArray destructor releases device memory automatically (RAII)
• Bidirectional data transfer — from_numpy() and to_numpy() for seamless CPU↔NPU exchange
• Broadcasting support — built-in GetBroadcastShape() follows NumPy broadcasting semantics
• Operator extensibility — missing CANN operators are supplemented by OpenBOAT

Installation

Requirements: GCC >= 11.2, CMake >= 3.26, Python >= 3.9, CANN >= 8.2.RC1.alpha003, Ascend 910B NPU.

Set the CANN environment variable before building:

export ASCEND_TOOLKIT_HOME=/usr/local/Ascend/ascend-toolkit/latest

Using uv (recommended)

git clone --recursive https://gitcode.com/cann/asnumpy.git
cd asnumpy
uv sync

Using pip

git clone --recursive https://gitcode.com/cann/asnumpy.git
cd asnumpy

# Upgrade pip, setuptools, and wheel to their latest versions
pip install --upgrade pip setuptools wheel

# Install the 'build' package, which provides a simple build front-end for Python packages
pip install build

# Build the current Python project (create source distribution and wheel)
python -m build

# Install all generated wheel (.whl) files from the dist/ directory
pip install dist/*.whl

Verify the installation

import asnumpy as ap
arr = ap.ones((1000, 1000), dtype=ap.float32)
print(arr.shape)  # (1000, 1000)

Quick Start

A minimal example is shown above. For a full side-by-side comparison of NumPy vs AsNumpy code and more usage patterns, see the Quick Start guide.

Runnable scripts are in examples/.

Performance

At 3000×3000 float32, ap.mean() runs 35.70× faster than np.mean() on the same machine. As data scale increases, the NPU's parallel computing advantage becomes more pronounced.

ap.mean() is a reduction operation; the top-section demo uses ap.multiply() (element-wise).

Shape	Data Size	AsNumpy (ms)	NumPy (ms)	Speedup
(500, 500)	250,000	0.1446	0.1429	0.99×
(1000, 1000)	1,000,000	0.1510	0.4904	3.25×
(2000, 2000)	4,000,000	0.1636	1.9372	11.84×
(3000, 3000)	9,000,000	0.1857	6.6303	35.70×

Full test environment, controlled variables, and reproduction instructions: benchmarks.md.

Roadmap

Release	Quarter	Key Deliverables
v0.3.0	26Q1	Documentation site (Docsify + GitCode Pages), CI/CD pipeline with hardware whitelist, code quality overhaul (spdlog, clang-format), PyPI release
v0.4.0	26Q2	Mathematical functions 100% API coverage, Ascend 950 validation, triton-ascend operator integration (10 ops), memory pool (experimental)
v0.5.0	26Q2	Linear algebra full coverage, triton-ascend operator library expanded to 20 ops, multi-NPU distributed computing research

Contributing

Contributions are welcome. Small fixes can be submitted directly as pull requests. For larger features, please open an issue first to discuss the design.

See the Developer Guide for build instructions, coding conventions, and how to add new operators.

Resources

• Documentation
• Quick Start
• Architecture
• Benchmarks
• FAQ
• Developer Guide
• Examples
• Issue Tracker
• OpenBOAT Operator Library

License

Apache License, Version 2.0 — see LICENSE.

AsNumpy is designed based on NumPy's API (see NumPy license).

AsNumpy is being developed and maintained by the School of Computer Science, Harbin Institute of Technology in deep collaboration with the Huawei CANN team, together with community contributors.

Acknowledgements

• AISS Group, School of Computer Science, Harbin Institute of Technology — Prof. Su Tonghua's team
• ISE Group, School of Computer Science, Harbin Institute of Technology — Prof. Wang Tiantian's team
• Huawei CANN team

---