Sparse-recovery decoder localizes 64 stuck cells in analog crossbars with 4-12 reads instead of 256
Connect_Agency_3975 · reddit · 2026-10-21
A preprint (code: NeuroFault on GitHub) shows a sparse-recovery decoder over random ±1 probe reads can localize up to 64 stuck cells in a simulated 256x256 analog crossbar tile with only 4-12 counted reads at 1% noise, versus 256 for full readout — 4.3-6.4x fewer reads than a hybrid baseline built on Wei & Roth's multi-outlier code (1.9-3.2x if noise-calibration reads are charged). A 16-parameter drift calibration fitted from few probe reads restores programmed accuracy on CIFAR-10, gaining +2.1 to +13.3 points over the best cheap digital baseline across 1h-1yr drift. Notably, the paper reports several pre-registered claims that failed, including transposed-read claims killed in follow-up (ratio 0.75 vs 0.7 bar at 2% noise), probe advantage limited to fault density ≤0.3%, and a refresh scheduler saving 9-13% of writes below the 15% bar. Limitations: simulation only, stuck-at faults only, possibly over-tight confidence bounds. The author requests critique on baselines and the noise model.
More from Infra
- OpenJet claims Qwen3.8-27B coding runs 1.9x faster than llama.cpp on one RTX 3090 — JeremyCMorgan · 2026-10-23
- Jensen Huang lays out NVIDIA's full Vera Rubin agentic system: NVL72, MGX, CMX and RTX — rohanpaul_ai · 2026-10-23
- ONNX to Transformers.js: running models directly in your browser — JosephJacks_ · 2026-10-23
- MiniMax M3.1 Flash free for 3 days on 0G, with 1M context and 512K output — MiniMax_AI · 2026-10-23
- "We bet on Cloudflare and it paid off": R2 and D1 at billion scale — dinasaur_404 · 2026-10-23
- Public data is drying up; private training data is turning into a browsable catalog — SucceededMind · 2026-10-23