Hi all,
We’re using the 32.2 image with the Jetson Nano.
Looking at the size of some of the libraries used, they seem really large.
Ex: libcudnn:
-rw-r--r-- 1 root root 383851496 feb 15 2019 /usr/lib/aarch64-linux-gnu/libcudnn.so.7.5.0
Is there some alternative image where the libraries are more reasonably sized (perhaps tailored only to work with the Nano?)
Regards,
Bogdan
Hi Bogdan,
As far as I know, there are no tailored versions of this libraries.
Depending on your case, you could use static linking.
That way, only those functions specific to your code should be included in your program.
I suspect you would still need to keep some of the content from /usr/lib/aarch64-linux-gnu/tegra and /usr/lib/aarch64-linux-gnu/tegra-egl.
Also, if you plan on using a csi camera you would still need to keep gstreamer and some other stuff around (no libv4l drivers for vi and isp and some dependencies on drm and x11).
At least, that’s what I’ve seen so far for the Nano.
If someone finds any mistake in my statements, I’d appreciate their corrections.
Best Regards,
Juan.
Hi again,
And thanks for the reply.
Static linking does not help since the binaries seems to runtime select what portion to use, i.e. all the code is referenced and can not be removed when linking anyway. (At least that’s my guess)
When looking at some of the binaries needed to run tensorRT stuff you’re looking at a minimum of more then 600MB(cublas, cudart, cudnn, nvinfer, etc). For those of use interested in basing a product on the Nano, it’s of course of interest to be able to OTA upgrade the device when needed and to keep the sizes of these upgrades as small as possible.
Anyone from Nvidia care to comment? Is there a plan to deliver nano-specific smaller libraries?
Regards,
Bogdan
Hi Bogdan,
I understand where you are coming from since I have worked on getting rid of that stuff.
So far, I haven’t seen anything of the likes, not for TK1, not for Nano, not for TX2.
My decision has been to avoid that stuff entirely, yet it can be a pain to rewrite everything from scratch.
It is also annoying to build an image with a size around 200MB of which 100MB are dependencies like libcuda (16MB), libnvidia-glcore (21MB) and libnvidia-eglcore (21MB).
I hope your question gets a more positive answer than I can provide.
Best Regards,
Juan.