We present an empirical method of evaluating high-speed serial link systems. Test cards implementing several hundred variations of the channel types expected in the real system are built; these channels are excited by the SerDes transceivers under consideration or an earlier version of the same. Performance is measured in terms of an available metric that is sufficiently dependent on the measure of interest; the available metric is empirically correlated with the measure of interest using calibration data. Link performance is parameterized against all variables of interest, and statistical analyses of performance and transceiver state matrices versus these variables are used to evaluate the solution at the system level.