Cool review! We have been looking at custom dataflow for floating-point computations -- http://dx.doi.org/10.1109/TCAD.2011.2173199 and http://dx.doi.org/10.1109/DFM.2014.21 but the big challenge has been handling dataflow graphs larger than chip capacity can hold.. Caching doesn't really work well as the dataflow wavefront proceeds along multiple separate paths.
