Demo report · ResNet-18 · public data · Jetson Orin NX
DEMO
Follow one image through the model: 0.850 ms as exported, 0.449 ms optimized.
As exported
reference
Optimized
−47.2 %
Latency measured on weights not yet retrained; accuracy known after retraining.
Slowed down 5,000 times · figures in µs
- 010.850 msAs exportedYour model as delivered.Reference
- 02 · In the pilot0.760 msOne change in your appThe model itself is untouched.
- 03 · In the pilot0.515 msOptimized, no retrainingThe check against your original model could not settle it.
- 04 · In integration0.449 msOptimized, with retrainingAccuracy known once retrained.
Measurement conditions
- As exported
- FP16 TensorRT engine built from the model as delivered. Every percentage is against it.
- One change in your app
- Same engine, launched through a CUDA graph.
- Optimized, no retraining
- INT8, calibrated on sample images, 2 calibration draws. Agreement with the original model’s predictions: 96.5 %, interval 94.9 to 97.6, against the 95 % demonstration floor. The interval straddles the floor, so the verdict is inconclusive.
- Optimized, with retraining
- Thinned channels, then INT8. Latency measured on weights not yet retrained, with provisional calibration. Accuracy is unknown until retraining.
- Per block
- Global pooling, +19 %: not judged against run-to-run noise.