I use Xavier NX for ZED 2 stereo camera. With same config, I got 30-35 FPS on 4.5 and 50-58 FPS on 4.4. Funny
Interesting - we using the PI High Quality Camera using the CIS interface - On 4.4 we were getting about 12 FPS. When we got the NX and installed 4.5.1 it throttles between 25 and 50 FPS but never settles down like the Nano. Our next step is move back to 4.4 on the NX and plots these numbers.
Same here, it seems that it impacts performance
I have to admit… i have reservations about Xavier NX performance… Nvidia is on the top of this issue ? Jetpack 4.4 vs 4.5.1
HW should provide maximum performance for us clients ?
we’re still investigate this issue internally, thanks
Can you give any updates how Nvidia is correcting performance issues ?
I know internal issues cant be said upfront but i guess we customers are eager to learn any positive corrective changes info we might get
Late to the party, however I’m getting the same throttling issue with the Xavier NX dev. kit running Jetpack 4.5 and the standard 19V 2.73A power brick. I saw this first while using the Stereolabs ZED SDK tools (ZED_Depth_Viewer & ZEDfu) while testing with a ZED Mini. I also didn’t see this issue with previous Jetpack versions.
Thanks for looking into this Nvidia, hopefully the internal investigations are going well.
Happened to me during visual recognition from camera, and also during ROS installation in 15w 6core mode. Previously did not happen.
could you please have the workaround to increase instantaneous OC limit, from 3.6A to 5A;
this 5A workaround should still acceptable with hardware aspect.
please have below command to set 5A instantaneous OC limit:
# echo 5000 > /sys/devices/c250000.i2c/i2c-7/7-0040/iio:device0/crit_current_limit_0
looking forward your feedbacks whether this approach works or not.
Device Info: Jetson Xavier NX (JetPack 4.5.0)
I changed the permissions, wrote 5000 and restore permissions
$ cd /sys/devices/c250000.i2c/i2c-7/7-0040/iio\:device0/ $ cat crit_current_limit_0 3600 ma $ sudo chmod 666 crit_current_limit_0 $ ls -l crit_current_limit_0 -rw-rw-rw- 1 root root 4096 Mar 25 16:10 crit_current_limit_0 $ sudo echo -n "5000" > crit_current_limit_0 $ sudo cat crit_current_limit_0 5000 ma $ sudo chmod 600 crit_current_limit_0 $ ls -l crit_current_limit_0 -rw------- 1 root root 4096 Mar 25 16:16 crit_current_limit_0 $ sudo cat crit_current_limit_0 5000 ma
I tried TensorFlow in Docker container. It looks worked!!!
Hi JerryChang, mehmetdeniz, worked for me too - no more throttling.
Hi again JerryChang, could you give a tip on how to make this workaround persistent - to work after a reboot?
Unfortunately for me i still keep getting warning (all 15W 2/4/6) modes and after acknowledging it keeps coming back
/sys/devices/c250000.i2c/i2c-7/7-0040/iio:device0$ sudo more crit_current_limit_0
Thanks that has worked for me.
you may have a workaround by updating
/etc/systemd/nv.sh, to include the command into initial scripts.
it looks some developers could workaround this by increasing instantaneous OC limit, from 3.6A to 5A;
suggest let’s having another new discussion thread to be created for issue tracking.
we’ll need to gather more details, and also check why it doesn’t works for all scenario.
Ran into this issue running yolov4–with stock Jetpack 4.5.1 (3600 mA limit) running yolov4-tiny would trip the alert. Once I raised the limit 5000 mA, I could run yolov4-tiny in 15W 6-Core with no alert, but running full-blown yolov4 would still trip it.
How can I do that please?
please check this discussion thread,
it’s confirmed this failure could be workaround by increased instantaneous OC limit from 3.6A to 5A.
however, it still able to reproduced with heavy GPU load even with 5A configuration.
let’s follow-up Topic 173571 for the use-case with heavy GPU load.