TensorRTPrecision
Enums · AiDotNet.Deployment.TensorRT
TensorRT precision modes for inference.
For Beginners
Precision affects speed vs accuracy trade-off:
- FP32 (32-bit float): Full precision, most accurate, slowest
- FP16 (16-bit float): Half precision, good balance, ~2x faster than FP32
- INT8 (8-bit integer): Quantized, fastest, requires calibration data
For most use cases, FP16 provides a good balance. Use INT8 only when you have calibration data and need maximum throughput.
Fields
| Field | Summary |
|---|---|
TensorRTPrecision FP16 |
16-bit floating point (half precision, ~2x faster) |
TensorRTPrecision FP32 |
32-bit floating point (full precision) |
TensorRTPrecision INT8 |
8-bit integer (quantized, fastest, requires calibration) |
int value__ |