Don't Read Every Cell: Few-Read Fault Localization and Drift Calibration in Analog Crossbars

Abstract

Deployed analog in-memory chips silently lose cells, and the default fix, reading every cell, costs 256 reads for a 256×256 tile. We show that a few random probe reads are enough. A sparse-recovery decoder over random ±1 probes localizes up to 64 stuck cells in a 256×256 tile with 4 to 12 counted reads at 1% read noise, on a simulated tile that is otherwise ideal. For k ≥ 4 faults it needs 4.3 to 6.4× fewer reads than a hybrid that we built from a published multi-outlier code (1.9 to 3.2× if its 8 noise-calibration reads are charged), and at least 2× fewer than tuned X-ABFT, which in this regime needs more reads than a full readout. The sharper finding is where the gain comes from: the decoder, not the second read direction. Forward reads alone already need 1.6 to 4.1× fewer reads than the hybrid; joint decoding with transposed reads needs a further 1.3 to 4× fewer at 1% noise, and gains less at higher noise. For drift, a 16-parameter field fitted per tile from a few probe reads restores programmed accuracy. As a digital correction that needs the stored weights, it gains +2.1 to +13.3 points over the best cheap digital baseline on CIFAR-10 from 1 h to 1 yr (+1.7 to +13.1 points with the 17 MVMs actually run); applied in the array, it keeps +0.6 to +3.7 points. Every claim is graded against criteria committed before the counted runs, and all code, data and pre-specified criteria are public.