GPU usage on K80 is much larger than Titan X

Memory usage on two machines are extremely, the one with K80 are using more than 2 times of the one with Titan X. Any thoughts? The system setups are the same: Ubuntu 16.04, cuda 8.0, cudnn 5.1. Thanks!


If cudnn.benchmark = True, it will automatically find the proper algorithms based on the GPUs, and may choose different convolution algorithms for K80 and TitanX, which have different memory consumptions. Maybe try to set it to False and see if they are the same.