How SK Hynix Builds Artificial Synapses for Brain-Like Computer Chips

A design for a tiny hardware component that mimics biological brain connections to help computers learn and process information like a human brain.

What it covers

This patent describes a physical component called a synapse designed for neuromorphic computing, which is a way of building computer chips that mimic the structure of the human brain. The core mechanism involves a 'reactive metal layer' shaped like a tapering wedge or staircase, positioned between two electrodes and an oxygen-containing layer. When voltage is applied, oxygen ions move to react with the metal, creating or removing a thin layer of insulating oxide. This change in the oxide layer alters the electrical conductivity of the device, effectively storing memory or 'weight' in the system, similar to how biological synapses strengthen or weaken their connections based on activity.

What it doesn't cover

• —Does not cover software-based neural networks that run on traditional CPUs or GPUs.
• —Does not cover synapses that rely on biological materials or organic chemistry.
• —Does not cover devices where the reactive metal layer has a uniform, non-tapered width.
• —Does not cover memory cells that do not use oxygen ion migration to change conductivity.

The clever bit

By tapering the width of the reactive metal layer, the device gains precise control over the area where the insulating oxide forms, which allows for more stable and predictable changes in electrical resistance.

Why it matters

Traditional computers are inefficient at tasks like pattern recognition because they separate memory from processing. Neuromorphic devices like this one aim to integrate memory directly into the processing unit, potentially allowing AI to run on tiny amounts of power. This is a critical step for moving advanced machine learning out of massive data centers and into local devices like phones or sensors.

Real-world examples

• 1.Experimental neuromorphic processor prototypes
• 2.Hardware-accelerated AI inference chips
• 3.Next-generation non-volatile memory research