AI BriefingSparse-recovery decoder localizes 64 stuck cells in 256x256 analog crossbar with 4-12 reads
·2026.10.21 19:09
Key point
Simulation shows a sparse-recovery decoder identifies up to 64 stuck cells in a 256x256 tile using 4 to 12 probe reads at 1% noise, compared to 256 reads for full readout.
Details
- Sparse-recovery decoder localizes up to 64 stuck cells in a 256x256 analog tile using 4 to 12 counted reads at 1% read noise, compared to 256 reads for full readout.
- For k >= 4 faults, the decoder uses 4.3 to 6.4 times fewer reads than a hybrid baseline built from a published multi-outlier code.
- Joint decoding with transposed reads requires a further 1.3 to 4 times fewer reads than forward reads alone at 1% noise, with gains diminishing at higher noise levels.
- A 16-parameter drift calibration restores programmed accuracy, gaining +2.1 to +13.3 percentage points over digital baselines on CIFAR-10 from 1 hour to 1 year of drift.
- Several pre-registered claims failed, including the transposed direction ratio at 2% noise (0.75 vs 0.7 bar) and drift structure-specificity (0.095 vs 0.15 gap).
- Results are simulation-only with stuck-at faults, white noise, and no silicon validation; confidence bounds are noted as too tight due to clustered instances.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
RFM requires GPUs simultaneously across three stages: training, simulation, and edge inference