Table of Contents
OCR Benchmark Update Overview
Document processing throughput directly impacts how quickly you can digitise archives, process invoices, or extract data from forms. This April 2026 benchmark update measures OCR speed across the leading open-source models on GigaGPU dedicated servers, using a standardised test set of 500 mixed document pages.
For the interactive benchmark tool, visit the OCR speed benchmarks page. This article highlights the key findings from the April 2026 test cycle.
Speed Results by Model and GPU
Pages per minute on mixed document types (invoices, letters, forms, tables):
| Model | RTX 3090 | RTX 5090 | RTX 5090 | RTX 6000 Pro |
|---|---|---|---|---|
| PaddleOCR v4 | 85 pg/min | 140 pg/min | 195 pg/min | 125 pg/min |
| Surya | 42 pg/min | 72 pg/min | 98 pg/min | 65 pg/min |
| GOT-OCR 2.0 | 35 pg/min | 58 pg/min | 82 pg/min | 52 pg/min |
| DocTR | 48 pg/min | 80 pg/min | 110 pg/min | 72 pg/min |
| EasyOCR | 55 pg/min | 92 pg/min | 125 pg/min | 82 pg/min |
PaddleOCR v4 remains the speed leader, processing 140 pages per minute on an RTX 5090. The RTX 5090 provides a meaningful 35-40% speed uplift across all models.
Accuracy vs Speed Trade-offs
| Model | Speed (RTX 5090) | Overall F1 | Table Accuracy | VRAM Used |
|---|---|---|---|---|
| PaddleOCR v4 | 140 pg/min | 93.5% | 89.8% | 2.8 GB |
| GOT-OCR 2.0 | 58 pg/min | 96.1% | 94.5% | 8.5 GB |
| Surya | 72 pg/min | 94.8% | 91.2% | 4.2 GB |
| DocTR | 80 pg/min | 92.0% | 87.3% | 3.5 GB |
GOT-OCR 2.0 delivers the highest accuracy but at the cost of speed and VRAM. For high-volume processing where 93%+ accuracy suffices, PaddleOCR v4 provides the best throughput. For accuracy-critical applications like financial documents, GOT-OCR 2.0 justifies the slower speed.
Batch Processing Throughput
Processing 10,000 pages end-to-end, including loading and post-processing:
| Model / GPU | Total Time | Effective Throughput |
|---|---|---|
| PaddleOCR v4 / RTX 5090 | 78 min | 128 pg/min |
| GOT-OCR 2.0 / RTX 5090 | 185 min | 54 pg/min |
| Surya / RTX 5090 | 148 min | 68 pg/min |
Batch throughput is slightly lower than peak per-page speed due to overhead from loading pages and writing results. For large-scale document processing projects, see the document processing throughput benchmark.
Cost per Page by Configuration
Based on GigaGPU hosting rates, running 24/7:
| Configuration | Pages/Month | Cost per Page |
|---|---|---|
| PaddleOCR / RTX 3090 ($175/mo) | 3.67M | $0.000048 |
| PaddleOCR / RTX 5090 ($250/mo) | 6.05M | $0.000041 |
| GOT-OCR 2.0 / RTX 5090 ($250/mo) | 2.33M | $0.000107 |
Self-hosted OCR costs a fraction of a cent per page, orders of magnitude cheaper than cloud OCR APIs. Detailed cost modelling available in the OCR cost per 10,000 pages guide.
Process Documents at Scale on Your Own Hardware
Self-hosted OCR with no per-page fees. Process millions of documents monthly on a dedicated GPU server.
View GPU ServersRecommendations
For maximum throughput on large document batches, PaddleOCR v4 on an RTX 5090 delivers the best pages-per-dollar. For complex documents requiring table extraction and layout understanding, GOT-OCR 2.0 justifies the additional cost. Both models leave substantial VRAM available for co-locating an open-source LLM for intelligent document processing pipelines.
Review the best OCR models guide for model selection and the GPU comparisons for hardware recommendations. Track the benchmarks section for future updates.
