On the Jetson TX1, looking at the cudaDeviceProp, the field totalGlobalMem shows that 2GB of the total unified 4GB are available to the GPU. Is it possible to get access to more than 2GB of memory for CUDA on the Jetson TX1?
It seems the totalGlobalMem device property reported by CUDA may not be completely accurate due to OS reasons.
In the past on TX1, I was able to allocate more memory than reported available (up to the amount indicated to be free by ~/tegrastats, or ~3.1GB).
It’s recommended to allocate smaller, individual chunks of buffers instead of trying to allocate it all at once.