optimized memory usage on the gpu: when used gpu_shared_var.value = ndarray, we keep only one memory region on the gpu.