The cuDNN and cuBlas libraries take up ~800MB GPU memory. I found in this post Why tensorRT occupy many memory ? , said that the library could be shared among all processes. But in my experiment, each process will own a separate copy. Is there any method to share those libraries?
May I know what’s the Jetson platform you’re using?
Which JetPack version?
I’m not using the Jetson platform. My hardware platform is NVIDIA 2080 Ti.
Hi @zykincs ,
Can you please help me with the cudnn version you are using?
I’m using the cudnn 8.04
Hi there, I met the same problem that each process owns a separate copy. Have you solved your problem?