nvidia

NVIDIA Jetson Thor

Blackwell-class humanoid robotics module. The headroom platform for transformer-heavy policies.

top models on this chip

Object detection

Sorted by latency p50 · matched-pair runs on NVIDIA Jetson Thor

Object detection

#	Model	Quant	Latency p50	Throughput	Acc.	Power
1	NVIDIA-YOLO27-Orin	MIXED	3.92 ms	255 FPS	99.1%	122.0 W
2	Meridian Highway-Perception v3	MIXED	4.09 ms	244 FPS	98.8%	91.6 W
3	RT-DETR-Edge	FP16	6.12 ms	163 FPS	99.9%	102.3 W
4	YOLO27-Edge	FP16	6.27 ms	160 FPS	99.6%	122.2 W
5	YOLO-World-Edge	FP16	6.60 ms	151 FPS	99.9%	100.2 W

top models on this chip

Sorted by latency p50 · matched-pair runs on NVIDIA Jetson Thor

Depth

#	Model	Quant	Latency p50	Throughput	Acc.	Power
1	Depth-Anything-Edge	MIXED	5.85 ms	171 FPS	99.3%	91.7 W

top models on this chip

Sorted by latency p50 · matched-pair runs on NVIDIA Jetson Thor

Segmentation

#	Model	Quant	Latency p50	Throughput	Acc.	Power
1	Mask2Former-Mobile	FP16	11.2 ms	89 FPS	99.5%	105.8 W
2	SAM-3-Distilled	FP16	13.1 ms	77 FPS	99.5%	96.4 W

top models on this chip

Sorted by latency p50 · matched-pair runs on NVIDIA Jetson Thor

Vision-Language-Action

#	Model	Quant	Latency p50	Throughput	Acc.	Power
1	Meridian LanePolicy-Mini	MIXED	25.0 ms	721 tok/s	99.2%	96.8 W
2	Cosmos-Policy-Lite	MIXED	29.0 ms	620 tok/s	98.7%	103.1 W
3	V-JEPA-AC-Edge	FP16	38.5 ms	467 tok/s	99.7%	94.2 W
4	RT-2-Mobile	MIXED	39.5 ms	455 tok/s	98.8%	113.6 W
5	OpenVLA-Edge	MIXED	56.7 ms	317 tok/s	99.0%	102.3 W

top models on this chip

Sorted by latency p50 · matched-pair runs on NVIDIA Jetson Thor

Multimodal

#	Model	Quant	Latency p50	Throughput	Acc.	Power
1	DINO-v3-Distilled	FP16	18.2 ms	55 FPS	99.8%	122.3 W
2	SmolVLM-3B	MIXED	110.8 ms	9 FPS	99.3%	111.0 W