speed for resnet-50_1by2 on gpu

[shesung]

As we know, the Flops of resnet-50_1by2 is about 1/4 of that resnet-50 has.

But I found that the trainning speed of resnet-50_1by2 on gpu is about 2/3 of resnet-50's. On cpu, the relation between the two speed is consistent with Flops.

the caffe is compiled with cuda v8.0 + cudnn v5

Anyone has some idea about this problem?