Toward Partitioned Reference Readout for Four-Level IGZO Memory: Device Statistics, Confusion Matrices, and INT4 Error Propagation

Abstract

Four-level indium-gallium-zinc oxide (IGZO) 2T0C memory needs reliable state formation, sensing references, and a path from cell errors to digital weights. We analyze 107 measured DC transfer curves from 37 isolated IGZO transistors (W/L = 3/0.5 micrometers). Four gate-bias operating points inferred from the first sweep are evaluated on later sweeps of 35 devices. Device-resolved operating points reduce second-sweep across-device current standard deviation by 49.4-52.4%, yet retaining initial thresholds increases misclassified points from 37/140 to 55/140. Leave-one-device-out reference estimation gives 0/140 on the second sweep; freezing second-sweep thresholds yields 6/140 on the third. Sampling eight reference transistors on the third sweep gives a mean held-out error fraction of 0.15% over 2,000 selections, versus 4.33% with frozen thresholds on matched held-out devices. We present conditional confusion matrices, a counterfactual common-gain sensitivity, and an INT4 error-propagation proxy. A simultaneous four-state bank with eight reference cells per state requires at least 32 physical reference cells, 12.5% of a proposed 256-cell data bank before peripheral overhead. A charge-conservation model and a partitioned hierarchical error model specify the pulse, retention, and array measurements needed to test transfer from DC device statistics to pulsed 2T0C operation.

Abstract: arXiv descriptive metadata (CC0). Thank you to arXiv for use of its open access interoperability. This service was not reviewed or approved by, nor does it necessarily express or reflect the policies or opinions of, arXiv.