A new AI semiconductor device independently manages long-term and short-term memory within a single chip, operating stably for over 10,000 seconds without performance degradation. Operating stably for over 10,000 seconds without performance degradation, the device marks a significant step towards reliable, on-device intelligence in neuromorphic computing hardware by 2026.
Yet, AI's explosive growth drives unsustainable energy consumption and reliance on external computing. Forbes projects electricity consumption from AI-optimized servers will rise nearly fivefold by 2030, creating an urgent demand for efficient, sustainable AI hardware.
Therefore, the industry is poised for a significant shift towards specialized, energy-efficient AI hardware. This could decentralize AI processing and enable new, previously impractical applications at the edge.
The Current Landscape: Large Systems vs. On-Device Needs
Intel's Hala Point neuromorphic research system, launched in 2024, simulates 1.15 billion neurons, according to Forbes, showcasing efforts in large-scale brain-inspired architectures. However, most flexible sensors and machine learning (ML) still rely on external computing for data analysis, causing high energy consumption, long latency, and privacy concerns, as noted by Nature. The contrast highlights the ongoing challenge of centralized, energy-intensive AI. New integrated devices aim to overcome this by moving intelligence closer to the data source.
A New Brain-Inspired Chip for Efficient AI
A research team developed a next-generation AI semiconductor device capable of independently implementing long-term and short-term memory within a single chip using surface acoustic waves (SAWs), reports Tech Xplore. The architecture directly addresses a core challenge in replicating brain-like memory functions in hardware.
The device uses electrical signals for long-term memory and SAWs for short-term memory, overcoming conventional memristor limitations in handling both functions simultaneously, Tech Xplore states. Its stable operation persisted for over 10,000 seconds without performance degradation.
