Easy benchmarking of all public open-source implementations of convnets. A summary is provided in the section below.
- Work in progress! After getting an initial baseline with the single module below (and getting inital benchmark scripts), I will benchmark a full AlexNet/MattNet/Overfeat *
Machine: 6-core Intel i7-3930K @ 3.20GHz + NVIDIA Titan Black + Ubuntu 14.04 x86_64
###Spatial Convolution layer (3D input 3D output)
#####Configuration: Input: 128x128 Batch-size 128, Feature maps: 3->96, Kernel Size 11x11
#####:forward()
| Library | Class/Function | Device | Input Config | GFlop/s | Code URL |
|---|---|---|---|---|---|
| cuda-convnet2 * | ConvLayer | GPU | DHWB | 1779.29 | Link |
| Caffe | ConvolutionLayer<Dtype> | GPU | BDHW | 1258.70 | Link |
| Torch-7 | nn.SpatialConvolutionMM | GPU | BDHW | 1177.78 | Link |
| ccv | ccv_convnet_layer | GPU | BDHW | 1024.16 | Link |
| cuda-convnet * | ConvLayer | GPU | DHWB | 929.17 | Link |
A * indicates that the library was tested with Torch bindings of the specific kernels.
Since this repository is getting a little attention, quickly adding some more results without making them pretty: cuda-convnet2 blows the competition out of the water by a huge margin!
-- layer 1
CONFIG: input = 3x128x128 * ker = 3x96x11x11 (bs = 128, stride = 1)
cuda-convnet: 944.68789645199 GFLOP/s (tm = 0.1314902305603)
Caffe/nn.SpatialConvolutionMM: 1177.7879871761 GFLOP/s (tm = 0.10546654462814)
cuda-convnet2: 1779.2908234941 GFLOP/s (tm = 0.069812774658203)
-- layer 2
CONFIG: input = 64x64x64 * ker = 64x128x9x9 (bs = 128, stride = 1)
cuda-convnet: 1260.782333477 GFLOP/s (tm = 0.42252349853516)
Caffe/nn.SpatialConvolutionMM: 2219.3736959591 GFLOP/s (tm = 0.24002724885941)
cuda-convnet2: 2197.8168036727 GFLOP/s (tm = 0.24238151311874)
-- layer 3
CONFIG: input = 128x32x32 * ker = 128x128x9x9 (bs = 128, stride = 1)
cuda-convnet: 1259.195897222 GFLOP/s (tm = 0.15540826320648)
Caffe/nn.SpatialConvolutionMM: 1164.72445476 GFLOP/s (tm = 0.16801351308823)
cuda-convnet2: 2244.3954637671 GFLOP/s (tm = 0.087190270423889)
-- layer 4
CONFIG: input = 128x16x16 * ker = 128x128x7x7 (bs = 128, stride = 1)
cuda-convnet: 1204.6621963356 GFLOP/s (tm = 0.017060458660126)
Caffe/nn.SpatialConvolutionMM: 490.99599360714 GFLOP/s (tm = 0.041857957839966)
cuda-convnet2: 2108.8329876002 GFLOP/s (tm = 0.009745717048645)
-- layers with small inputs/kernels, seen at the lower ends of the network
CONFIG: input = 384x13x13 * ker = 384x384x3x3 (bs = 128, stride = 1)
cuda-convnet: 983.54876028249 GFLOP/s (tm = 0.041795969009399)
Caffe/nn.SpatialConvolutionMM: 735.82546939461 GFLOP/s (tm = 0.05586701631546)
cuda-convnet2: 2283.9565044269 GFLOP/s (tm = 0.01799875497818)