PaddleOCR (Baidu) combines detection, recognition, and structured layout analysis. The 5060 Ti hosts it at high throughput for document-AI workflows.
TL;DR
PaddleOCR on a 5060 Ti processes ~12 A4 pages/sec for plain text, ~6 pages/sec for layout-aware extraction (PP-StructureV2). Plenty for a contract-processing or invoice-parsing pipeline.
Setup
- PaddleOCR 2.7
- PP-OCRv4 detection + recognition
- Optional: PP-StructureV2 for layout
- FP16 inference
Results
| Workload | Pages/sec |
|---|---|
| Plain text A4 | ~12 |
| Plain text + table extraction | ~9 |
| Layout-aware (PP-StructureV2) | ~6 |
| Multi-language (CN/EN/JP) | ~9 |
| Receipt parsing | ~25 |
Verdict
For document-AI deployments at small-medium scale (thousands of pages/day), the 5060 Ti is the cost leader. Pair with an LLM for structured extraction.
Bottom line
PaddleOCR on a 5060 Ti + Llama 3.1 8B for post-processing = complete document-AI pipeline at £119/mo.
