Scientists Build AI Chip That Cuts Calculations 10,000x as Arrhythmias Are Flagged at 98% Accuracy
Updated
Updated · Livescience.com · Aug 11
Scientists Build AI Chip That Cuts Calculations 10,000x as Arrhythmias Are Flagged at 98% Accuracy
2 articles · Updated · Livescience.com · Aug 11
Summary
Simulated tests showed the cerebellum-inspired chip detected ECG arrhythmias within one-fifth of a heartbeat, hitting 98% accuracy and running more than twice as fast as a standard transformer model.
The gain comes from a memtransistor that combines memory and processing in one component, easing the von Neumann bottleneck and switching between excitatory and inhibitory modes to ignore routine signals.
Researchers modeled the design on the cerebellum rather than the cerebrum, aiming to reserve computing power for unexpected events instead of continuously analyzing every data stream.
The approach could support low-power edge AI in health monitors, robots, autonomous vehicles and cybersecurity systems, potentially reducing dependence on energy-hungry data centers whose electricity demand could reach 945 terawatt-hours by 2030.
The team reported the results July 10 in Nature Communications and said the next step is teaching the hardware to adapt as repeated events become predictable.