TL;DR In simulation, a sparse-recovery decoder over a few random +/-1 probe reads localizes up to 64 stuck cells in a 256x256 analog tile with 4 to 12 counted reads at 1% read noise. Reading every cell takes 256 reads. The paper also separates the decoder's gain from the second read direction's gain, and gives a 16-parameter drift calibration. Several pre-registered claims failed, and the paper reports them.

Setting. An analog crossbar tile does matrix-vector products inside a memory array. A stuck cell stays at one value. A probe read sends a random +/-1 input vector through the tile and reads the output. Sparse recovery means we assume few cells are bad and search for the short list of cells that explains the reads.…