1,200 images/sec OCR

TurboOCR is an OCR engine that reimplements PaddleOCR with C++/CUDA, FP16 TensorRT, kernel fusion, batch recognition, and a multi-stream pipeline to dramatically boost speed.

• It accepts images and PDFs over HTTP/gRPC and returns bounding boxes, text, layout regions.
• Layout recognition can be toggled on or off per request, and enabling it only adds about 20% to inference time.
• It claims performance of 270 img/s on text-heavy pages and 1,200+ img/s on sparse pages.
• The target environment is Linux, RTX 50-series, CUDA 13.2.

The author explains that the project was created because VLM/OCR cost and throughput became a bottleneck in bulk PDF processing. However, they note that for complex table extraction or structured output like invoice → JSON, VLM-based OCR such as PaddleOCR-VL is still more suitable.

In other words, this is a notable open-source optimization case for workloads where OCR throughput matters, such as large-scale document indexing or real-time RAG.