Neuromorphic AI Processor/IP

Memory and Computing, Reimagined

Conventional AI processors perform computations by repeatedly moving large amounts of data between memory and the computing unit.

MEWLINK’s ReRAM-CiM Neuromorphic Core is based on a Computing-in-Memory architecture that stores AI weights in ReRAM and performs core neural network operations within memory.

This reduces the burden of data movement during AI computation and enables a neuromorphic computing structure for high-efficiency AI inference.

Key Technologies

ReRAM Crossbar Architecture

A CiM structure that stores neural network weights and performs parallel computations by configuring ReRAM cells in a crossbar shape.

2T2R ReRAM Cell

Supports stable weight representation and neuromorphic computation based on a symmetric ReRAM cell structure.

In-Memory AI Computing

Combines memory and computation to minimize data movement and efficiently process core neural network operations.

Compute Only When It Matters

Traditional frame-based AI iteratively processes the entire data regardless of whether the input data changes.

MEWLINK’s Neuromorphic AI Processor IP utilizes Event-Driven SNN Processing to pursue a structure that selectively performs computations when a valid event occurs.

Through event-centric sparse processing, it reduces unnecessary computations and supports efficient AI inference in real-time Edge AI environments where fast response times are required.

Event → Spike → SNN Processing → AI Inference

Processing only the necessary information at the necessary moment is the core of the neuromorphic AI computation pursued by MEWLINK.

Toward High-Precision Neuromorphic Computing

In ReRAM-based analog computation, Device-to-Device and Cycle-to-Cycle Variations resulting from the physical characteristics of the device can affect the stability and accuracy of AI computations.

MEWLINK is developing technology that reduces the impact of such physical deviations in ReRAM devices on computations and enables stable weighted computations through Bipolar Signal Control and Self-Calibration technologies.

By combining symmetric control of bipolar signals with ReRAM characteristic correction technology, we aim for more stable and precise neuromorphic AI computations.

Bipolar Signal Control → Variation Compensation → Self-Calibration → Stable AI Computing

From Neuromorphic Core to AI Processor

MEWLINK is expanding the ReRAM-CiM Neuromorphic Core into an AI Acceleration Processor IP capable of integrating with RISC-V-based processor architectures, rather than keeping it as an independent computational block.

By integrating the RISC-V Processor, Neuromorphic AI Accelerator, Memory Controller, and System Interface, it aims for an integrated SoC architecture capable of supporting everything from AI inference to system control.

RISC-V Processor

• System Control & Management

Neuromorphic AI Accelerator

• ReRAM-CiM & Event-Driven SNN Processing

Memory Architecture

• On-Chip / Off-Chip Memory Integration

System Interface

• Scalable interface between AI Accelerator and various systems

From Core IP to Edge AI

The Neuromorphic AI Processor envisioned by MEWLINK is not a dedicated processor for a single specific application.

We are building a scalable AI Processor IP Platform capable of integration into various Edge AI systems, centered on ReRAM-CiM-based computation and Event-Driven SNN Processing.

We aim to expand into an IP-based technology platform that links the Neuromorphic Core with the AI Acceleration Architecture according to customers' AI models, sensor interfaces, and system requirements.

Target Applications

Vision AI

Real-time video and event-based AI processing

Robotics

Intelligent Robotics Requiring Low-Latency Recognition and Judgment

Autonomous Systems

Autonomous AI requiring real-time environment perception and decision-making

Drones & Mobility

Mobile Edge Devices Require Low Power and Real-Time AI Processing

Intelligent Sensors

Intelligent Edge combining sensors and AI processing

Edge AI & IoT

On-Device AI Processing with Reduced Cloud Reliance

Memory. Computing. Intelligence.

MEWLINK develops a new AI processor architecture that breaks down the boundaries between memory and computing and combines event-driven intelligence.

It reduces data movement with ReRAM-CiM,

Performs computations at the moment they are needed with event-driven SNNs,

Aims for stable, high-precision computation through bipolar signal control and self-calibration, Expands to various AI systems through RISC-V architecture.

ReRAM-CiM × Event-Driven SNN × RISC-V

Building the Next Generation of Neuromorphic AI Processor IP