Compute Command Encoder Question

In the documentation it says "Batch 8 or more compute commands into the same MTLComputeCommandEncoder." How is this done? I was under the impression that individual kernels were encoded in single command encoders using `set_compute_pipeline_state` and the inputs / outputs were set that way, and then many command encoders were created on the same buffer. Is it possible to do multiple kernels per command encoder, and is there a benifit over one kernel per command encoder, many command encoders per command buffer?