Dual-domain architecture shows 39.2X higher energy efficiency running artificial neural networks inferencing with floating-point dataflow

https://techxplore.com/news/2025-02-dual-domain-architecture-higher-energy.html

"Analog Computing-in-Memory processes/ stores data in single hardware... progress has mostly been in classification tasks (e.g., images), not in handling complex regression (e.g., You Only Look Once) object detection), due to high computational noise/ floating-point data compatibility issues... overcome combining high-precision floating point compatible digital computing/ energy efficiency... prototype memristor-based system attains 2.7X higher average precision than that of pure Analog Computing-in-Memory systems"

Related:

Novel memristors to overcome AI's 'catastrophic forgetting'
https://techxplore.com/news/2025-03-memristors-ai-catastrophic.html#google_vignette

Analog computer processes/ stores information in same location, similar to biological neurons; steers robot using 0.25% of the power needed by conventional controllers
https://news.engin.umich.edu/2025/03/brain-like-computer-steers-rolling-robot-with-0-25-of-the-power-needed-by-conventional-controllers/

Critical band-to-band-tunnelling based optoelectronic memory
https://www.eurekalert.org/news-releases/1076491