My summary: * deephi built an FPGA based processing unit specialized for sparse low precision tensor operations * Now they have a dev kit for turning your ten

My summary: * deephi built an FPGA based processing unit specialized for sparse low precision tensor operations * 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. They published a paper about how they balance compression/accuracy. [0] [0] https://arxiv.org/pdf/1612.00694.pdf https://arxiv.org/pdf/1612.00694.pdf