CVCUDA/CV-CUDA

CV‑CUDA – GPU‑accelerated computer‑vision library

What it is

• An open‑source NVIDIA library that implements a large set of image‑ and video‑processing operators (resize, color conversion, filters, etc.) that run entirely on the GPU.
• Designed to be a high‑throughput, low‑latency building block for AI pipelines on NVIDIA hardware – from cloud servers to Jetson edge devices.

Key capabilities

• GPU‑native operators – All functions operate on GPU memory, avoiding costly host‑device transfers.
• Python and C/C++ APIs – Use the library from Python (via cvcuda wheels) or directly from C/C++ code.
• Broad hardware support – Works on x86_64 and aarch64 (including ARM‑based Grace and Jetson) with CUDA 12.x or 13.x.
• Integration with other NVIDIA tools – Example shown with nvImageCodec for GPU‑direct JPEG decode, and the library can be chained with other CUDA‑accelerated components.
• Scalable packaging – Pre‑built Python wheels, Debian packages, and source builds; only one CUDA version can be installed at a time.

Typical use‑cases

• Pre‑processing frames for deep‑learning inference (e.g., resizing, color conversion, cropping) at video‑scale rates.
• Real‑time visual‑search pipelines where latency matters.
• Edge AI on Jetson devices that need to process camera streams without CPU bottlenecks.
• Any computer‑vision workflow that benefits from keeping data on the GPU throughout the pipeline.

Getting started

# Install the wheel that matches your CUDA version
pip install cvcuda-cu12   # for CUDA 12.x
# or
pip install cvcuda-cu13   # for CUDA 13.x

A minimal Python example (shown in the README):

import cvcuda
from nvidia import nvimgcodec

decoder = nvimgcodec.Decoder()
img = decoder.read('input.jpg')          # GPU‑direct decode
tensor = cvcuda.as_tensor(img, 'HWC')    # wrap as CV‑CUDA tensor
resized = cvcuda.resize(tensor, (224,224,3), cvcuda.Interp.LINEAR)

For full installation details—including building from source, Docker dev images, and platform‑specific notes—see the online docs: https://cvcuda.github.io/CV-CUDA/installation.html.

Project health & community

• Actively maintained by NVIDIA (latest version v0.19.0).
• Pre‑built binaries for Linux (many‑linux, Ubuntu 22.04+, WSL2) and aarch64; Windows support only via WSL2.
• Contributions are currently limited to NVIDIA internal teams, but users can file bugs or feature requests through GitHub issues.
• Licensed under Apache‑2.0, allowing commercial use and modification.

Why it matters CV‑CUDA lets developers keep the entire vision preprocessing stack on the GPU, eliminating the classic CPU‑GPU bottleneck that limits throughput in large‑scale AI services. By providing a drop‑in, GPU‑accelerated alternative to CPU‑based libraries (e.g., OpenCV), it helps AI teams achieve higher frame rates and lower latency for tasks such as visual search, video enhancement, and real‑time analytics.