CV-CUDA™ is an open-source library that enables building high-performance, GPU-accelerated pre- and post-processing for AI computer vision applications in the cloud at reduced cost and energy. It provides a specialized set of 45+ highly performant computer vision and image processing operators, making it ideal for various AI imaging and computer vision workloads deployed at scale in the cloud.

Features

• Specialized Kernels: Offers a specialized set of 45+ highly performant computer vision and image processing operators.
• API Support: Provides C, C++, and Python APIs for flexible integration.
• Batching Support: Supports batching with variable shape images.
• Zero-Copy Interfaces: Integrates seamlessly with deep learning frameworks like PyTorch and TensorFlow.
• Inference Server Example: Includes an NVIDIA Triton™ Inference Server example using CV-CUDA and NVIDIA® TensorRT™.
• End-to-End Acceleration: Provides end-to-end GPU-accelerated object detection, segmentation, and classification examples.

Benefits

• Cost and Energy Efficiency: Hand-optimized kernels save cost and energy, making it suitable for cloud-based use cases.
• High Throughput: Achieves up to 49X end-to-end throughput improvement, moving bottlenecked pre- and post-processing pipelines from the CPU to the GPU.
• Scalability: Enables developers of cloud-scale applications to save tens to hundreds of millions in compute costs and eliminate thousands of tons in carbon emissions.
• Interoperability: Works with libraries, SDKs, and frameworks such as nvJPEG, Video Codec, Video Processing Framework (VPF), TAO Toolkit, TensorRT, Triton Inference Server, PyTorch, and TensorFlow.