The Achilles' heel of memristive technologies (www.nature.com)

A new review argues that the biggest practical weakness of memristive devices—used for nonvolatile memory and in-memory/neuromorphic compute—is not just materials physics but poor and inconsistent testing practices. Lanza et al. show that claims about endurance, retention and system-level performance are frequently overstated because studies use non‑standardized protocols, small sample sizes, selective statistics, and unrealistic simulations. That has produced scatter in the literature, hindered reproducibility, and slowed commercialization by creating mismatched expectations between academia and industry. For the AI/ML community the consequences are concrete: memristor-based accelerators and on‑chip neural networks depend on device-level reliability (cycle-to-cycle variability, retention loss, forming voltages, temperature sensitivity) and on accurate system-level models. The authors call for community standards (e.g., JEDEC-like test suites), robust statistical reporting, realistic endurance/retention testing, full-stack co‑simulation, and design-for-variability measures (error correction, redundancy, algorithmic tolerance). They also note routes to mitigate problems via better materials/process control and honest benchmarking. Adopting standardized characterization and transparent statistics will be essential for trustworthy memristive hardware in ML systems.