Why is transposed convolution faster ?

[EvergreenTree]

This is the result of 2D convolution

ConvNd time: 6209.3720703125
ConvGT time: 13.088768005371094
convND error: 3.196690840923111e-07 %

ConvTransposeNd time: 32.824127197265625
ConvTransposeGT time: 21.486591339111328
convTransposeND error: 4.2337585881568884e-08 %

and this is the result of 5D convolution

ConvNd time: 1249.503173828125
torch.Size([1, 1, 2, 2, 2, 2, 2])
ConvTransposeNd time: 39.06355285644531
torch.Size([1, 1, 11, 11, 11, 11, 11])

I wonder why the transposed convolution is comparable to built-in pytorch method but convolution is much slower?