Performance Summary

The performance results of benchmarking the Isaac ROS packages on both Jetson and x86_64 platforms are below:

Node	Input Size	AGX Thor T5000	AGX Thor T4000	DGX Spark	x86_64 w/ RTX 5090
AprilTag Node	720p	385 fps 2.9 ms @ 30Hz	280 fps 11 ms @ 30Hz	462 fps 2.4 ms @ 30Hz	596 fps 2.0 ms @ 30Hz
FoundationPose Pose Estimation Node	720p	3.15 fps 370 ms @ 30Hz	3.40 fps 350 ms @ 30Hz	2.77 fps 360 ms @ 30Hz	10.1 fps 99 ms @ 30Hz
DNN Stereo Disparity Node Full	576p	175 fps 8.8 ms @ 30Hz	160 fps 26 ms @ 30Hz	167 fps 8.2 ms @ 30Hz	350 fps 5.3 ms @ 30Hz
DNN Stereo Disparity Node Light	288p	350 fps 3.5 ms @ 30Hz	349 fps 12 ms @ 30Hz	350 fps 4.1 ms @ 30Hz	350 fps 3.5 ms @ 30Hz
Stereo Disparity Node	1080p	217 fps 5.7 ms @ 30Hz	196 fps 20 ms @ 30Hz	249 fps 5.4 ms @ 30Hz	747 fps 4.0 ms @ 30Hz
Rectify Node	1080p	1550 fps 1.6 ms @ 30Hz	768 fps 4.8 ms @ 30Hz	2500 fps 2.3 ms @ 30Hz	2500 fps 2.9 ms @ 30Hz
TensorRT Node DOPE	VGA	165 fps 6.4 ms @ 30Hz	147 fps 27 ms @ 30Hz	107 fps 8.6 ms @ 30Hz	300 fps 4.0 ms @ 30Hz
Triton Node DOPE	VGA	161 fps 7.5 ms @ 30Hz	153 fps 30 ms @ 30Hz	94.5 fps 10 ms @ 30Hz	281 fps 4.4 ms @ 30Hz
TensorRT Node PeopleSemSegNet	544p	566 fps 2.1 ms @ 30Hz	407 fps 6.8 ms @ 30Hz	594 fps 1.4 ms @ 30Hz	1570 fps 1.5 ms @ 30Hz
Triton Node PeopleSemSegNet	544p	361 fps 4.0 ms @ 30Hz	280 fps 11 ms @ 30Hz	362 fps 3.1 ms @ 30Hz	485 fps 5.4 ms @ 30Hz
DNN Image Encoder Node	VGA	356 fps 11 ms @ 30Hz	300 fps 19 ms @ 30Hz	318 fps 10 ms @ 30Hz	417 fps 10 ms @ 30Hz
Occupancy Grid Localizer Node	~50 sq. m	28.1 fps 40 ms @ 30Hz	21.1 fps 110 ms @ 30Hz	40.1 fps 27 ms @ 30Hz	50.1 fps 13 ms @ 30Hz
H.264 Decoder Node	1080p	296 fps 4.5 ms @ 30Hz	296 fps 5.3 ms @ 30Hz	736 fps 4.4 ms @ 30Hz	296 fps 4.1 ms @ 30Hz
H.264 Encoder Node I-frame Support	1080p	296 fps 2.9 ms @ 30Hz	297 fps 10 ms @ 30Hz	921 fps 2.7 ms @ 30Hz	1050 fps 3.8 ms @ 30Hz
H.264 Encoder Node P-frame Support	1080p	296 fps 2.9 ms @ 30Hz	297 fps 9.5 ms @ 30Hz	846 fps 3.1 ms @ 30Hz	1050 fps 3.8 ms @ 30Hz
Nvblox Node	–	4.94 fps 89.8 ms	–	–	4.96 fps 23.1 ms

Graph	Input Size	AGX Thor T5000	AGX Thor T4000	DGX Spark	x86_64 w/ RTX 5090
AprilTag Graph	720p	354 fps 5.2 ms @ 30Hz	273 fps 15 ms @ 30Hz	546 fps 3.9 ms @ 30Hz	596 fps 2.8 ms @ 30Hz
Centerpose Pose Estimation Graph	VGA	50.2 fps 27 ms @ 30Hz	50.4 fps 73 ms @ 30Hz	50.2 fps 23 ms @ 30Hz	50.2 fps 19 ms @ 30Hz
DOPE Pose Estimation Graph	VGA	143 fps 22 ms @ 30Hz	116 fps 53 ms @ 30Hz	89.3 fps 27 ms @ 30Hz	193 fps 17 ms @ 30Hz
DNN Stereo Disparity Graph Full	576p	157 fps 13 ms @ 30Hz	68.5 fps 38 ms @ 30Hz	150 fps 13 ms @ 30Hz	347 fps 9.9 ms @ 30Hz
DNN Stereo Disparity Graph Light	288p	285 fps 9.2 ms @ 30Hz	155 fps 20 ms @ 30Hz	231 fps 11 ms @ 30Hz	349 fps 9.2 ms @ 30Hz
Stereo Disparity Graph	1080p	203 fps 10 ms @ 30Hz	185 fps 26 ms @ 30Hz	183 fps 9.0 ms @ 30Hz	600 fps 6.9 ms @ 30Hz
DetectNet Object Detection Graph	544p	143 fps 18 ms @ 30Hz	111 fps 43 ms @ 30Hz	88.0 fps 19 ms @ 30Hz	227 fps 18 ms @ 30Hz
Grounding DINO Object Detection Graph	544p	23.4 fps 50 ms @ 30Hz	22.3 fps 90 ms @ 30Hz	17.5 fps 63 ms @ 30Hz	130 fps 15 ms @ 30Hz
RT-DETR Object Detection Graph SyntheticaDETR	720p	188 fps 12 ms @ 30Hz	156 fps 35 ms @ 30Hz	165 fps 12 ms @ 30Hz	444 fps 11 ms @ 30Hz
TensorRT Graph PeopleSemSegNet	544p	449 fps 8.1 ms @ 30Hz	319 fps 19 ms @ 30Hz	562 fps 6.7 ms @ 30Hz	1330 fps 6.3 ms @ 30Hz
SAM Image Segmentation Graph Full SAM	720p	2.26 fps 350 ms @ 30Hz	2.24 fps 290 ms @ 30Hz	2.22 fps 280 ms @ 30Hz	20.8 fps 57 ms @ 30Hz
SAM Image Segmentation Graph Mobile SAM	720p	15.0 fps 230 ms @ 30Hz	15.0 fps 200 ms @ 30Hz	14.6 fps 82 ms @ 30Hz	70.3 fps 20 ms @ 30Hz

Methodology

The performance numbers provided above are measured using Isaac ROS Benchmark. Configuration files for reproducing results are provided in the launch scripts for each benchmark here.

Frames per second (fps) are measured using the maximum sustained framerate for the accelerated computational pipeline on each listed platforms. Performance measured includes input node → graph of nodes under performance test → output node where the publishing rate of the input node is auto-tuned to discover the peak throughput dropping <5% of the frames. Average fps computed over 5 runs, discarding minimum and maximum runs; latency measured at 30hz publishing rate matching sensor capture rate.