1 Reply Latest reply on Feb 2, 2018 7:47 AM by noctis

    Threadripper 1950x bad performance for tensorflow

    noctis

      Hi, guys,

       

      I recently tested some deep learning applications including inception V3, Resnet implemented with TensorFlow on my machine. The machine has 2 1080ti and 1950x. And the memory speed is 2933 with  64GB capacity.

       

      I saw from some websites that one 1080ti with intel CPU can deal with ~140 pictures per second, while mine can only deal with 80+ pictures. With 2 GPU running together, that figure increases to 150 pictures per second. This means that two 1080ti with 1950x can only reach the performance of 1 1080ti on Intel platform!!!! This is too terrible!

       

      Do you guys have any related experience? I searched the internet but found no solution. There's even no bazel build option for AMD cpu.

       

      Looking for your help. I'm in despair , for spending twice the money to buy such a machine with so disappointing performance.