-
1 BSP environment:
TX2 jetpack 4.6 L4T R32.6.1 kernel 4.9 aarch64
TX2 (p3310)
-
2 Background:
we recompiled kernel, and used flash.sh to update the system.img to the target device tx2. now, we need cuda to accelerate computing, but how to install it?
-
3 Problem:
3.1 according to jetpack 4.6, jetpack-sdk-46, it includes cuda 10.2. how to use jetpack to install cuda 10.2.
3.2 is it possible to local install cuda 10.2 manually?
FYI
ubuntu@ubuntu-desktop:~$ ls /usr/local
bin etc games include lib man sbin share src
ubuntu@ubuntu-desktop:~$ cat /etc/nv_tegra_release
# R32 (release), REVISION: 6.1, GCID: 27863751, BOARD: t186ref, EABI: aarch64, DATE: Mon Jul 26 19:36:31 UTC 2021
ubuntu@ubuntu-desktop:~$ lspci
ubuntu@ubuntu-desktop:~$ nvidia-smi
-bash: nvidia-smi: command not found
ubuntu@ubuntu-desktop:~$ nvcc -v
-bash: nvcc: command not found
Hi Henry,
The way to download CUDA is through the skdmanager. After you open it, on Step 2, you can select so that it only installs the SDK Components but doesn’t do the flash.
Alternatively, you can follow this guide :
https://developer.ridgerun.com/wiki/index.php/Installing_CUDA-toolkit_and_VisionWorks_Jetson_TX2/Xavier/Nano
You’ll still need to download the SDK components using the sdkmanager (Just by selecting Download now. Install later.) But the actual installation can be done manually.
Hi
on step 2, we ticked the “jetson SDK components” box only, dis-ticked the “jetson OS” box. Then, it comes to the step 3. what does the ipv4 address mean? what should we do next step?
FYI
nvsdkm.tar.gz (488.2 KB)
I do not know, where does the “tsinghua.edu.cn” come from, since the source.list does not contain it.
Hi
now, i can use the sdkmanager to install the “target components”, such as cuda, cuDNN, tensorRT etc.
but, it is not convenient to install cuda at mass production stage. we expect to install cuda in the source code, so that we can compile it in the firmware, such as system.img, and flash the fw to target devices.
how?
Hi,
You need to press Download now. Install later in the first screenshot that you sent, then you can follow the link I provided.
For a mass production stage, massflash is probably what you are looking for:
https://docs.nvidia.com/jetson/archives/r34.1/DeveloperGuide/text/SD/FlashingSupport.html?highlight=massflash%20sh#flashing-to-multiple-jetson-devices
With this, you can set up a single device as you need it to be (including any cuda, drivers, etc you want). And using this method you’ll have a massflash binary that will flash any other device in such a way that it reproduces the setup for any other device after it has been flashed.
