A Jetson Nano Ubuntu 20.04 with OpenCV, TensorFlow and Pytorch BUT OVERCLOCKED?

[srevandro]

Dear @Qengineering,

How are you?

Thanks for Jetson Nano Ubuntu 20.04 image. I would like to know with you don´t have the Ubuntu 20.04 with OpenCV, Tensorflow, Pytorch etc, but overclocked? I just saw here the Bare Image (without any ML library) that is overclocked.

I have tried to your tutorial to overclock the Ubuntu 20.04 with all libraries, but did not succeed. I just get some errors:

make[2]: [scripts/Makefile.build:649: drivers/net/wireless] Error 2 make[2]: Waiting for unfinished jobs.... CC [M] drivers/net/veth.o CC [M] drivers/net/macvlan.o CC [M] drivers/net/macvtap.o make[1]: [scripts/Makefile.build:649: drivers/net] Error 2 make: [Makefile:1122: drivers] Error 2

I have looked at the NVIDIA forum, but did not succeed as well. https://forums.developer.nvidia.com/t/overclocking-jetson-nanos-cpu-and-gpu/83501/11

Do you have any idea how to solve this problem? I have tried to copy, but did not work. sudo cp /proc/config.gz $TEGRA_KERNEL_OUT/

Another possibility is do you have any tutorial about the installation of OpenCV, Tensorflow, Pytorch in the Bare Image? I think should be easier.

Thanks.

[Qengineering]

I've chosen not to overclock the Nano because some frameworks may become unstable. Not that it will happen, but chances are. Other frameworks don't profit from overclocking due to their use of CUDA. The second reason is that the files involved in overclocking are frequently updated by NVIDIA. After a "sudo apt-get upgrade" or the software updater, overclocking can be undone. On the other hand, apart from these remarks, there is no reason overclocking won't work. The procedure is always the same. Get a fresh Tegra OS, overclock and then upgrade to Ubuntu 20. If you want to install frameworks on the bare (and overclocked) image, please use our guides on the website. We also use them to compose this SD image.

[srevandro]

Dear @Qengineering ,

I have done some attemps to get Jetson nano overclock. Following I will list my procedures and problems that I get:

1) I have got you Ubuntu 20.04 LTS FULL (with Pytorch, OpenCV etc), and I tried to overclock it. It does not work. I got the previous errors mentioned before. make[2]: *** [scripts/Makefile.build:649: drivers/net/wireless] Error 2 make[2]: *** Waiting for unfinished jobs.... CC [M] drivers/net/veth.o CC [M] drivers/net/macvlan.o CC [M] drivers/net/macvtap.o make[1]: *** [scripts/Makefile.build:649: drivers/net] Error 2 make: *** [Makefile:1122: drivers] Error 2

2) I get the original sd card image from NVIDIA with Ubuntu 18.04 LTS (https://developer.nvidia.com/embedded/learn/get-started-jetson-nano-devkit#write), and tried to overclock it. On this case I got now drivers missings errors and succeed on the overclock. However the big problem in this version is that they use Python 3.6.9, and some libraries (such as matplotlib pillow etc) requires at this Python 3.7.

3) I get you Bare Image already overclocked. and following your tutorials on your site I succeed in install Pytorch. However it seems GPU(CUDA) is not working on this case. So my first question if I ONLY need to do the procedures in the "Installation by wheel" or should I do any other to get GPU(CUDA) working?

4) Do you think I also will have problems with GPU (CUDA) in the installation of torchvision, OpenCV, Tensorflow ?

5) Last question, is a silly one, but I need to be sure I am doing the correct thing. Suppose I install you Bare Image with overclocked system, and later on I install Jetson-Inference container and Jetson-Containers, and try to run the libraries torch, torchvision etc from this containers. My question if this containers will use the CPU (1.9GHz) and GPU (0.95GHz) overclocked or not since they are containers?

Thanks for all you support, and I will be looking forward your support. It´s really important to me succeed on this issue, because I really need an overclocked system with all deep learning libraries install.

Sincerely yours, Evandro

[Qengineering]

1) It is possible the install OpenCV, TensorFlow and PyTorch on the bare overclocked image. I've tested it with PyTorch. Below the result.

2) I'm uploading the SD-card to Sync. I will share the link as soon as the upload is completed. It holds PyTorch. Not OpenCV nor TensorFlow. If you don't mind, you can give it a try.

3) At the moment you upgrade your system (or your clients) you will lost the overclocking. Don't use sudo apt-get upgrade.

4) Not sure if the Jetson Containers are holding the complete OS. (Seems to be??)

[srevandro]

Dear @Qengineering ,

I am really really grateful for all you help.

I was a little bit confused with NVIDIA instructions to update the pytorch version, since they recommend to update a version where it was using Python3.6. About the CUDA is working perfectly. This was my own mistake.

I have tried now install NVIDIA 18.04 LTS, update to Ubuntu 20.04 LTS, and applied overclock instructions. At the final kernel rebuild I got the problems before related to wireless drivers and another one, but seems to be OK.

I will try to execute the deep learning algorithms and verify which one can be ran in the overclock mode.

I am writing an article about this, and I would like to add my thanks to you for all the help. Should I just reference your website or do you want to add your full name ?

Thanks.

[Qengineering]

1) sudo apt-get upgrade isn't the problem. The problem arise as soon as the upgrade procedure wants to (partly) replays the kernel. That's were things go wrong. A regular upgrade of FireFox, for instance, is no problem.

2) The /usr/bin/jetson_clocks --show reflects the values in the governor, the values wherein the clock may vary. If it says 1900, you know your Jetson Nano is still overclocked.

3) Please find an already overclocked Ubuntu 20.04 version with PyTorch. The one giving the screen dump from above. No need to rebuild the kernel. That's done. You only have to install OpenCV en TensorFlow yourself. The Image is about 7 GB. https://ln5.sync.com/dl/778d47210/pkg4aht5-7mt4hxb3-k4t2pc4g-zz6rxjwh

4) Reference to the website would be appreciated.

[srevandro]

Thanks my friend. You are the best. Regards.