Run PaddleOCR on a Dedicated GPU Server

GPU Selection for PaddleOCR

PaddleOCR is PaddlePaddle’s open-source OCR toolkit supporting text detection, recognition, and layout analysis in 80+ languages. It is remarkably lightweight, making even budget GPUs viable for high-throughput PaddleOCR hosting on a dedicated GPU server:

PaddleOCR uses under 1 GB for standard text recognition, meaning you can easily co-host it alongside an LLM like Phi-3 or Qwen 2.5 on the same GPU for document understanding pipelines.

Install PaddleOCR

# Install PaddlePaddle GPU and PaddleOCR
pip install paddlepaddle-gpu paddleocr

# Basic OCR usage
from paddleocr import PaddleOCR

ocr = PaddleOCR(use_angle_cls=True, lang="en", use_gpu=True)
result = ocr.ocr("document.png", cls=True)

for line in result[0]:
    coords, (text, confidence) = line
    print(f"[{confidence:.2f}] {text}")

PaddleOCR auto-downloads model weights on first run. It supports English, Chinese, Japanese, Korean, and 80+ other languages out of the box.

Building an OCR API

# Install FastAPI
pip install fastapi uvicorn python-multipart

# api.py
from fastapi import FastAPI, UploadFile
from paddleocr import PaddleOCR
import tempfile, os

app = FastAPI()
ocr = PaddleOCR(use_angle_cls=True, lang="en", use_gpu=True)

@app.post("/ocr")
async def extract_text(file: UploadFile):
    with tempfile.NamedTemporaryFile(delete=False, suffix=".png") as tmp:
        tmp.write(await file.read())
        tmp_path = tmp.name
    result = ocr.ocr(tmp_path, cls=True)
    lines = []
    for line in result[0]:
        coords, (text, confidence) = line
        lines.append({"text": text, "confidence": float(confidence)})
    os.unlink(tmp_path)
    return {"lines": lines}

# Run: uvicorn api:app --host 0.0.0.0 --port 8000

Read the self-host guide for server setup fundamentals. Check the OCR speed benchmarks for cross-GPU performance data.

Performance Benchmarks

Tested with A4 scanned documents at 300 DPI (2480×3508 pixels).

The RTX 4060 processes 200 pages per minute with the standard pipeline, making it ideal for high-volume document digitisation. The RTX 3090 pushes this to 300 pages per minute for enterprise-scale workloads.

Optimisation Tips

• Batch multiple pages to keep the GPU fully utilised during sequential document processing.
• Use TensorRT acceleration for a 2-3x throughput improvement over the default PaddlePaddle backend.
• Pre-process images to consistent DPI and orientation before OCR to improve accuracy and speed.
• Use PP-OCRv4 for speed and PP-Structure only when you need table extraction or layout analysis.
• Co-host with an LLM to build intelligent document processing pipelines that extract and summarise text in a single pass.

Compare GPU options with the GPU comparisons tool. For cost planning, use the cheapest GPU for AI inference guide.

Next Steps

PaddleOCR is one of the most efficient AI workloads to self-host. Pair it with Whisper for multi-modal document and audio processing. For text analysis after extraction, see our LLaMA hosting options. Browse all deployment guides in the model guides section.