They (including the 1080ti which is basically a Titan) do support 4x faster INT8, though, so if comparing to a reduced-precision ternary net running a FPGA, tha

They (including the 1080ti which is basically a Titan) do support 4x faster INT8, though, so if comparing to a reduced-precision ternary net running a FPGA, that seems relevant. (They mention using INT8 in some of the GPU benchmarks but I'm not sure which graphs are supposed to represent that.)