Benchmarks for popular convolutional neural network models on CPU and different GPUs, with and without cuDNN.

Some general conclusions from this benchmarking:

• Pascal Titan X > GTX 1080: Across all models, the Pascal Titan X is 1.31x to 1.43x faster than the GTX 1080 and 1.47x to 1.60x faster than the Maxwell Titan X. This is without a doubt the best card you can get for deep learning right now.
• GTX 1080 > Maxwell Titan X: Across all models, the GTX 1080 is 1.10x to 1.15x faster than the Maxwell Titan X.
• ResNet > VGG: ResNet-50 is faster than VGG-16 and more accurate than VGG-19 (7.02 vs 9.0); ResNet-101 is about the same speed as VGG-19 but much more accurate than VGG-16 (6.21 vs 9.0).
• Always use cuDNN: On the Pascal Titan X, cuDNN is 2.2x to 3.0x faster than nn; on the GTX 1080, cuDNN is 2.0x to 2.8x faster than nn; on the Maxwell Titan X, cuDNN is 2.2x to 3.0x faster than nn.
• GPUs are critical: The Pascal Titan X with cuDNN is 49x to 74x faster than dual Xeon E5-2630 v3 CPUs.

All benchmarks were run in Torch. The GTX 1080 and Maxwell Titan X benchmarks were run on a machine with dual Intel Xeon E5-2630 v3 processors (8 cores each plus hyperthreading means 32 threads) and 64GB RAM running Ubuntu 14.04 with the CUDA 8.0 Release Candidate. The Pascal Titan X benchmarks were run on a machine with an Intel Core i5-6500 CPU and 16GB RAM running Ubuntu 16.04 with the CUDA 8.0 Release Candidate. The GTX 1080 Ti benchmarks were run on a machine with an Intel Core i7-7700 CPU and 64GB RAM running Ubuntu 16.04 with the CUDA 8.0 release.

We benchmark all models with a minibatch size of 16 and an image size of 224 x 224; this allows direct comparisons between models, and allows all but the ResNet-200 model to run on the GTX 1080, which has only 8GB of memory.

The following models are benchmarked:

Top-1 and Top-5 error are single-crop error rates on the ILSVRC 2012 Validation set, except for VGG-16 and VGG-19 which instead use dense prediction on a 256x256 image. This gives the VGG models a slight advantage, but I was unable to find single-crop error rates for these models. All models perform better when using more than one crop at test-time.

Speed is the total time for a forward and backward pass on a Pascal Titan X with cuDNN 5.1.

You can download the model files used for benchmarking here (2.1 GB); these were converted from Caffe or Torch checkpoints using the convert_model.lua script.

We use the following GPUs for benchmarking:

(input 16 x 3 x 224 x 224)

We use the BVLC AlexNet from Caffe.

AlexNet uses grouped convolutions; this was a strategy to allow model parallelism over two GTX 580 GPUs, which had only 3GB of memory each. Grouped convolutions are no longer commonly used, and are not even implemented by the torch/nn backend; therefore we can only benchmark AlexNet using cuDNN.

(input 16 x 3 x 224 x 224)

We use the Torch implementation of Inception-V1 from soumith/inception.torch.

(input 16 x 3 x 224 x 224)

This is Model D in [3] used in the ILSVRC-2014 competition, available here.

(input 16 x 3 x 224 x 224)

This is Model E in [3] used in the ILSVRC-2014 competition, available here.

(input 16 x 3 x 224 x 224)

This is the 18-layer model described in [4] and implemented in fb.resnet.torch.

(input 16 x 3 x 224 x 224)

This is the 34-layer model described in [4] and implemented in fb.resnet.torch.

(input 16 x 3 x 224 x 224)

This is the 50-layer model described in [4] and implemented in fb.resnet.torch.

(input 16 x 3 x 224 x 224)

This is the 101-layer model described in [4] and implemented in fb.resnet.torch.

(input 16 x 3 x 224 x 224)

This is the 152-layer model described in [4] and implemented in fb.resnet.torch.

(input 16 x 3 x 224 x 224)

This is the 200-layer model described in [5] and implemented in fb.resnet.torch.

Even with a batch size of 16, the 8GB GTX 1080 did not have enough memory to run the model.