In my experience, the fully connected layers are the bottleneck. The other issue was the alternating compute-heavy convolution and the IO-heavy pooling. I'm curious how this FFT implementation stacks up against cuDNN (what's the speedup like for just the convolutional layers? and then what's the overall speedup like?).
Comments
In my experience, the fully connected layers are the bottleneck. The other issue was the alternating compute-heavy convolution and the IO-heavy pooling. I'm curious how this FFT implementation stacks up against cuDNN (what's the speedup like for just the convolutional layers? and then what's the overall speedup like?).
http://arxiv.org/pdf/1412.7580v2.pdf compares the convolutional implementation with the cuDNN layers. For the FC layers, it's just CuBLAS `sgemm`.