Skip to content Skip to main menu

Products

Neuromorphic AI Processor/IP

Neuromorphic AI Processor/IP

Next-generation Neuromorphic Processor IP for ultra-low power, ultra-high speed, and ultra-precision AI computing

Neuromorphic AI Acceleration Processor/IP

Next-generation Neuromorphic Processor IP for ultra-low power, ultra-high speed, and ultra-precision AI computing

Next-generation AI Acceleration Processor IP combining ReRAM-based Computing-in-Memory (CiM) technology and Spiking Neural Network (SNN)-based event-driven computation technology.

Reduces data movement and unnecessary computations that occur in existing processor-centric AI computation structures, and enables ultra-low power, ultra-high-speed AI computation optimized for edge environments through a neuromorphic architecture that selectively performs computations when necessary events occur.

IP Innovation Technology

Core Technology

ReRAM-CiM-based AI Acceleration

Minimizes data movement by placing memory and computation close together, efficiently performing core AI neural network operations within memory.

Event-Driven SNN Processing

Reduces unnecessary computations by applying an event-signal-based selective computation structure, providing a low-latency processing environment suitable for real-time AI inference.

High-Precision Neuromorphic Computing

Supports high-precision AI computation by compensating for physical deviations in ReRAM devices through simultaneous control of +/- bipolar signals and the application of self-calibration technology.

RISC-V-based SoC Integration

Applicable to various Edge AI systems based on an SoC architecture capable of integrating neuromorphic accelerators, RISC-V processors, memory, and interfaces.

AI Acceleration Processor IP

AI Acceleration for the Edge

AI Acceleration Processor IP is scalable into a next-generation computing platform for Vision AI, Robotics, Drones, Intelligent Sensors, IoT, and various real-time Edge AI systems.

• Ultra-Low Power

• Ultra-High Speed

• Ultra-Precision

ReRAM Hybrid-CiM Neuromorphic Core

A neuromorphic computing core that stores AI model weights in ReRAM, performs neural network computations directly in memory without continuously moving data out of memory, and processes the results through SNN-based event computations.

In the AI Acceleration Processor IP, the ReRAM Hybrid-CiM Neuromorphic Core is the core acceleration computing engine responsible for actual AI computation.

ReRAM Hybrid-CiM Neuromorphic Core

Why use ReRAM for "AI computations"?

Simplifying the basic operations of an AI neural network, it is:

Input × Weight → Accumulation → Result.

Weights are particularly important in neural networks.

ReRAM can store these weights by utilizing the resistance state of each cell.

Why use ReRAM for "AI computations"?

The Key is the "ReRAM Crossbar"

The most important structure of the ReRAM Hybrid-CiM Core is the Crossbar Array.

A ReRAM Cell exists at each intersection point.

Therefore, a structure can be created where:

• Row → Input or signal

• Column → Calculation result reflecting weights

• ReRAM Cell → Weight storage and participation in calculations

For this reason, it is suitable for processing Matrix-Vector Multiplication (MVM) of many neural networks in parallel on hardware.

The Key is the "ReRAM Crossbar"

Why the "2T2R" Structure?

Conceptually, 2T2R utilizes two transistors and two ReRAMs as a single memory cell structure.

The key point is not simply "using ReRAM," but designing a stable 2T2R-based CiM Cell by considering the physical characteristics of ReRAM and the asymmetry and variability that may occur in surrounding circuits.

Why the "2T2R" Structure?

Bipolar Signal Simultaneous Control Self-Calibration

By symmetrically utilizing positive (+) and negative (-) signals, it compensates for physical deviations in ReRAM that occur commonly and reduces the effects of asymmetry, thereby processing neural network weights more stably.

Bipolar Signal Simultaneous Control Self-Calibration

SNN Engine

ReRAM-CiM is a hardware engine that performs computations efficiently, while SNN is a neuromorphic computation method that performs computations only at the moment an event occurs.

ReRAM-CiM + Event-Driven SNN → Ultra-low power, ultra-high-speed computation

SNN Engine