The easiest pipeline optimisation, which is basically free hardware-wise and is documented in the academic literature for years, is to move the computation of W + K + H to the previous pipeline stage, reducing the critical path. This is a trivial win because all the data is available already and H can be discarded afterwards, which also has the nice bonus of mapping more efficiently to certain FPGA hardware. It sounds like Intel have taken this and expanded on it to move even more things into earlier and later stages of the pipeline.

Or, as Intel puts it, "The precomputed (H.sub.i+K.sub.i+W.sub.i) may be stored in the 32-bit register 402 dedicated for H.sub.i. This optimization reduces the critical path for the computation of E.sub.i+1 by one CSA or approximately three logic gates." I implemented this back in 2011 in some open source FPGA mining code, and it was an old trick even back then. They don't seem to be citing any prior art, which is a bit dubious.