[SOLVED] Titan V on PyTorch 0.3.0, CUDA 9.0, CUDNN 7.0 is much slower than 1080 Ti

vision
5

With torch.backends.cudnn.benchmark = True added, here are the results!!!

vgg16 passes:

  • 1080 Ti: 41.4ms
  • Titan V: 31.3ms

resnet152 passes:

  • 1080 Ti: 60.4ms
  • Titan V: 49.0ms

densenet121 passes:

  • 1080 Ti: 29.9ms
  • Titan V: 26.2ms

Looks like adding this magic line works!!!

Thanks @Soumith_Chintala!
As a user, I was expecting that simply popping in a Titan V would work a little faster than a 1080 Ti without modifying my code (I posted a toy example here, but my VGG16-based MVCNN that I used for object classification ran much slower as well, so that’s why I posted this.)

1 Like