Confirmed. If add "with tf.device('/cpu:0'):" to the program I listed, it runs much faster (like 10 times faster than on GPU), and faster than on my Intel MBP.