This seems to get impressive results cross compiling models from torch (with vollo-torch) into FPGA friendly formats. Speedups look quite impressive...