Experiments on torch1.60 with 2080ti show that disabling cuDNN results in a significant training speedup, however it is uncertain whether this affects performance. However, it's very strange and no idea about why cause it.
Top is set torch.backends.cudnn.enabled = False.
Under is default Settings.
.
Experiments on torch1.60 with 2080ti show that disabling cuDNN results in a significant training speedup, however it is uncertain whether this affects performance. However, it's very strange and no idea about why cause it. Top is set torch.backends.cudnn.enabled = False. Under is default Settings.
.