gpuarray 0.3.0

NumPy-like GPU array library that works on every GPU — NVIDIA, AMD, Intel, Apple Silicon. No CUDA required.

gpuarray

NumPy-like GPU array library that works on every GPU — NVIDIA, AMD, Intel, Apple Silicon. No CUDA required.

Powered by WebGPU via wgpu-py. Uses Metal on macOS, Vulkan on Windows/Linux, and DirectX 12 on Windows.

Install

pip install gpuarray

Usage

import gpuarray as gp

# Create arrays on GPU
a = gp.array([1, 2, 3, 4, 5])
b = gp.ones(5) * 3

# Operations run on GPU
c = a + b
d = a * b
e = gp.dot(a, b)

# Read back to CPU
print(c.to_numpy())  # [4. 5. 6. 7. 8.]

# Matrix multiply
A = gp.ones((512, 512))
B = gp.ones((512, 512))
C = gp.matmul(A, B)  # runs on GPU

# Reductions
total = gp.sum(a)
avg = gp.mean(a)

# Math functions
x = gp.exp(a)
y = gp.log(a)
z = gp.sqrt(a)

Why?

CuPy only works on NVIDIA GPUs with CUDA. gpuarray works on every GPU:

GPU	CuPy	gpuarray
NVIDIA (CUDA)	✓	✓
AMD (Vulkan)	✗	✓
Intel (Vulkan)	✗	✓
Apple Silicon (Metal)	✗	✓

Supported Operations

Array Creation

array, zeros, ones, arange, linspace

Elementwise Binary

+, -, *, /, ** (with arrays or scalars)

Elementwise Unary

exp, log, sqrt, abs, neg, relu, sigmoid, tanh

Reductions

sum, mean, max, min

Linear Algebra

dot, matmul

Performance

Benchmarks on Intel UHD 630 (integrated GPU) vs NumPy (CPU with SIMD):

Operation	NumPy	gpuarray	Speedup
add 10M elements	11.6ms	19.9ms	0.59x
matmul 512x512	1.3ms	1.0ms	1.22x
sum 10M	4.3ms	12.5ms	0.34x

The integrated GPU shows speedup on compute-heavy operations (matmul). Discrete GPUs (RTX 3080, RX 7900, etc.) will show much larger speedups across all operations.

Requirements

• Python 3.10-3.13
• Any GPU supported by WebGPU (most GPUs from 2015+)
• No CUDA, no special drivers — just your system GPU

How It Works

1. Arrays are stored as WebGPU GPU buffers
2. Operations dispatch WGSL compute shaders to the GPU
3. Pipelines are cached — repeated operations don't recompile
4. Results stay on GPU until you call .to_numpy()

License

MIT