Error-aware Probabilistic Update: Memristor-based hardware train AI models with 6 orders of magnitude better energy efficiency than GPUs

https://techxplore.com/news/2026-01-memristor-method-slashes-ai-energy.html

"hardware-algorithm mismatch hinders analog in-memory computing for training deep neural networks: difficult to precisely write to/ backpropagation requires tiny, deterministic weight adjustments but memristor noise/ drift creates update errors 10 to 100X larger than intended adjustment... overcome embracing noise using probabilistic scaling precise updates converted into larger discrete pulses applied with specific probability... skips update if smaller than noise threshold... reduces writes by over 99%"