Inference Benchmarks

Measured throughput for RF-DETR, YOLOv8, ResNet, and ViT models on NVIDIA L4 and Jetson Orin NX, comparing ONNX and TensorRT.

This page contains performance benchmarks for various models running with Inference on different hardware platforms.

NVIDIA L4 GPU

Object detection

Performance benchmarks for object detection models, comparing standard ONNX runtime with TensorRT-optimized adapters:

Model Type	Size	Inference/sec (ONNX)	Inference/sec (TRT)	Improvement Multiplier
rfdetr-nano	384x384	103.5	299.5	2.9
rfdetr-small	512x512	57.4	253.4	4.4
rfdetr-medium	576x576	62.8	201.8	3.2
rfdetr-large	704x704	36.9	160.5	4.3
rfdetr-xlarge	700x700	18.1	96.1	5.3
rfdetr-2xlarge	880x880	17.4	74.1	4.3

Segmentation

Performance benchmarks for instance segmentation models:

Model Type	Size	Inference/sec (ONNX)	Inference/sec (TRT)	Improvement Multiplier
rfdetr-seg-nano	312x312	51.9	105.3	2.0
rfdetr-seg-small	384x384	57.5	126.7	2.2
rfdetr-seg-medium	432x432	39.7	99.7	2.5
rfdetr-seg-large	504x504	32.8	93.2	2.8
rfdetr-seg-xlarge	624x624	17	68.8	4.0
rfdetr-seg-2xlarge	768x768	10.7	59	5.5

Classification

Performance benchmarks for classification models:

Model Type	Size	Inference/sec (ONNX)	Inference/sec (TRT)	Improvement Multiplier
ResNet50	224x224	358.6	600.8	1.7
ViT	224x224	238	306.5	1.3

Jetson Orin NX

Object detection

Performance benchmarks for object detection models on Jetson Orin NX:

Model Type	Size	Inference/sec (ONNX)	Inference/sec (TRT)	Improvement Multiplier
rfdetr-nano	384x384	21.2	78.5	3.7
rfdetr-small	512x512	13.9	52.5	3.8
rfdetr-medium	576x576	11	44	4.0
yolov8n-640	640x640	35.6	89	2.5
yolov8s-640	640x640	26.3	69.5	2.6
yolov8m-640	640x640	13.5	44.5	3.3
yolov8l-640	640x640	9	32.5	3.6
yolov8x-640	640x640	6.4	22	3.4

Benchmark methodology

All benchmarks were conducted using the Inference CLI, with a single image (batch size 1, -bs 1) over 500 iterations (-bi 500).

• Inference/sec (ONNX) - standard ONNX runtime, measured with:

  inference benchmark python-package-speed -m [model]

• Inference/sec (TRT) - TensorRT-optimized adapters (supported in inference 1.0 and later), measured with:

  USE_INFERENCE_MODELS=TRUE inference benchmark python-package-speed -m [model]