How hard to run YOLOv10x on Orin Nano

Robotics & Edge Computing Jetson & Embedded Systems Jetson Orin Nano
1

Hi, I’m very interested in how much resources YOLOv10x requires on Orin Nano. As I understand Ultralytics (YOLOv10 - Ultralytics YOLO Docs) v10-X requires 160 GFLOPS to run a 640(times 480, I assume) video with 10.7 ms latency, executing on COCO dataset.
A Jetson Nano 8 GB delivers 40 TFLOPS @ 15 W. Does this mean that running this YOLO application on Nano theoretically requires much less than a thousandth of its calculation capacity, and some fractions of 1 W of its power?
Best regards, Torbjörn Martinsson

3

Hi,

15W is a system power consumption rather than just AI engines.
So it should be that YOLOv10 can reach a certain fps under 15W power consumption.

Thanks.

4

Thank you for your reply. I fully understand that 15 W is system power, and that I can run YOLOv10 on a Nano. But my interest is in how much power Nano allocates to do that - with 10.7 ms latency. As Nano has 40 TFLOPS capability, I imagine it may run another application requiring 39.84 TFLOPS, at the same time as running a YOLOv10x application - at 15 W in total system power. And if so, this means that Nano can run this YOLO with much less than 1 % of its total computational resources, while utilizing much less than 1 W of its 15 W total power.
Best regards, Torbjörn

5

Hi,

40TFLOPS is a theoretical number but 10.7 ms is measured in practice.
So maybe you can try to test the maximum throughput of YOLOv10 (ex. batch) and then compute the expected power consumption of a single batch to get the number.

Thanks.

6

Maybe I should have explained that I’m working on a paper where I compare computers to human brain. I’m investigating several options available today, and it seems like Nano has a very interesting performance. Though, I haven’t consider buying one and do testing, but instead ask someone with insight in this matter. Is there maybe another place at Nvidia where I could get some estimates of this?
Best regards, Torbjörn

7

Hi,

Do you want some data for computation per watt?
If yes, we have something similar in DLA which might help:

Please note that the data required extra hardware to measure the thermal data.
The hardware is only for internal usage and is not public.

Thanks.

8

Thank you for the update. If I understood the link accurately, it was about a relative efficiency comparing DLA with GPU’s - not absolute values. I hope I got it right.

9

Hi,

Unfortunately, we don’t have the absolute value.

Thanks.

11

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.