Performance Tips

Performance Tips

English | 日本語

Real-time processing cost is determined by the entire chain, not only by the name of one filter.

Most effective ways to reduce load

1. Lower capture resolution.
   This is often the most effective change because ONNX cost rises quickly with pixel count.

2. Use lighter ONNX models.
   Model architecture matters more than file size alone.

3. Reduce interpolation multiplier.
   x5 requires more generated frames than x2.

4. Remove redundant filters.
   Avoid multiple resize, sharpen, or restoration passes unless they visibly help.

5. Use duplicate-frame reduction when appropriate.

6. Use an FPS limit when processing more source frames is unnecessary.

7. Try TensorRT on supported NVIDIA RTX GPUs.

GPU utilization can be misleading

A low percentage in Windows Task Manager does not necessarily mean the GPU has large usable performance headroom.

Modern GPUs dynamically change core and memory clocks. A light shader may show the same percentage as a heavier workload while running at much lower clocks.

Different GPU engines can also be reported separately.

High-refresh output

To actually display 120 fps, the complete path must sustain approximately 8.33 ms per presented frame.

For 240 fps, the available frame time is approximately 4.17 ms.

A filter that takes longer than the target frame interval cannot maintain that output rate by itself.

Measure the whole chain

Compare:

• Capture only
• GLSL only
• One ONNX stage
• Interpolation only
• Your complete intended preset

This makes it much easier to find the actual bottleneck.