My summary:<p>* deephi built an FPGA based processing unit specialized for sparse low precision tensor operations<p>* Now they have a dev kit for turning your tensorflow/caffe/mxnet models into compressed (sparse/low precision) binaries that are optimized to run on their processors.<p>They published a paper about how they balance compression/accuracy. [0]<p>[0] <a href="https://arxiv.org/pdf/1612.00694.pdf" rel="nofollow">https://arxiv.org/pdf/1612.00694.pdf</a>