I have some code that works fine in the Orin Nano 8GB devkit. When I run it in the production board, though, I get an “out of memory error” if I try to user cudaMallocHost, even for 1 byte (error code number 2). cudaMallocManaged fails as well. There’s plenty of RAM free in the system, and straight malloc seems to work.
I also get this error message in the terminal one or more times, seemingly before that:
NvMapMemAllocInternalTagged: 1075072515 error 12
NvMapMemHandleAlloc: error 0
My code uses nvimgcodec to save an image in jpeg.
How can I even debug this? What kinds of test could I do relative to memory?
Both systems are running the same CUDA and driver versions.
UPDATE: rebooting seems to help.