Stride is the amount by which a kernel "slides" in a convolution. The "default" stride is 1 -- which usually goes unmentioned in publications. Bigger strides have the effect of downsampling the input -- so they can be used to replace the combination of a stride-1 convolution followed by pooling.