Extended Data Fig. 5: Weight programming procedure.
a, Crossbar array during programming. b, Proposed TDP algorithm to program a target conductance value G on a unit-cell. c, Weight error comparison between TDP of this work and previous approaches. TDP - Max-fill refers to programming the two devices with iterative programming up to the ODP \({G}_{\max }\), as proposed in Ref. 31. Due to the wide SET distribution shown in Fig. 2a, some devices in the core either cannot achieve \({G}_{\max }\), or conversely could be programmed to much higher conductance values than \({G}_{\max }\). Therefore, the latter approach leads to programming inaccuracies resulting from either under-utilizing the conductance range of individual devices or from devices that cannot reach \({G}_{\max }\). The proposed TDP algorithm solves this issue by using the readout SET conductance of the devices of the unit-cell to map the weight.
