High Performance Implementation of 3D Convolutional Neural Networks on a GPU

Table 3

Performance of cuDNN SGEMM versus that of the 3D WMFA on 3D convolution layers. Performance is measured in effective TFLOPS.

Layer	× × × ×		TFLOPS		Speedup
			cuDNN SGEMM	3D WMFA
conv2	32 × 16 × 56 × 56 × 32	64	1.21	1.28	1.05
conv3	64 × 8 × 28 × 28 × 32	256	2.38	3.31	1.39
conv4	256 × 4 × 14 × 14 × 32	256	2.4	4.72	1.96
conv5	256 × 2 × 7 × 7 × 32	256	1.46	2.1	1.44