Intel Loihi 2 Neuromorphic Processor : Specifications, Architecture, Working, Differences & Its Applications

The latest generation of neuromorphic processors like Loihi 2 is the successor to Loihi, introduced by Intel Labs in late 2021. It is a second-generation neuromorphic research test chip that utilizes an asynchronous SNN (spiking neural network) to execute adaptive, self-improving parallel computations, thereby implementing inference & learning with high efficiency. Intel Loihi2 aims to achieve scales of neural complexity by drawing inspiration from neuroscience to inform its architectural stimulation. This processor improves its predecessor to develop the computational capacities & efficiency of silicon neuromorphic systems mainly for real-time intelligent processing. This article elaborates on the Intel Loihi2 Neuromorphic Processor, its working, and its applications.

What is Intel Loihi2 Neuromorphic Processor?

Loihi 2 is the latest neuromorphic research chip by Intel, which implements spiking neural networks through modular connectivity, programmable dynamics, and optimizations for speed, efficiency, and scale. This processor includes six Lakemont microcontrollers & 128 completely asynchronous neuron cores, which are connected by a network-on-chip.

The neuron cores can be optimized for neuromorphic workloads by executing a set of spiking neurons with all connected synapses to such neurons. Thus, the communication between all the neuron cores will be in the form of a spike message. Likewise, Microprocessor cores can also be optimized for executing standard C code and spike-based communication to help with data I/O, network configuration, management & monitoring.

The Intel Loihi2 has some new functionality for implementing custom neuron models with microcode instructions. Thus, generates and transmits graded spikes by supporting three-factor learning rules. A single Loihi 2 chip supports up to 120 million synapses and 1 million neurons.

Please refer to this link to know more about Brainchip Akida.

How does the Intel Loihi2 Neuromorphic Processor Work?

The second-generation neuromorphic processor, like Intel Loihi 2, works by imitating how the human mind can process data through asynchronous events, spikes, and an extremely parallel architecture. The breakdown of = Loihi 2 processor working is discussed below.

Spiking Neural Networks

Loihi 2 neuromorphic processor processes data with spikes (discrete events) instead of traditional neural networks. Neuron spikes within Loihi 2 fire whenever their internal state crosses a threshold. Thus, it leads to event-driven and asynchronous processing similar to how actual neurons communicate.

Architecture based on Parallel and Mesh

Loihi 2 neuromorphic processor includes up to 1 million neurons & 120 million synapses within a single chip. Thus, it utilizes a mesh network to route discrete events between cores efficiently, where every core has many neuro-cores that replicate neuron & synapse populations.

Asynchronous and Clockless Design

Loihi 2 processor asynchronously works without a global clock to reduce power consumption greatly by allowing for real-time responsiveness. Whenever spikes occur, events are processed only by allowing ultra-low power operation.

Plasticity & On-Chip Learning

This processor supports STDP local learning rules by allowing on-chip learning without a CPU or cloud. So, this is key for edge AI applications, wherever adaptability is required without any external training.

LAVA Programmability

The Lava software framework is developed by Intel to program the Loihi 2 processor. So Lava allows developers to write neuromorphic applications within Python to run them across traditional and neuromorphic hardware.

Specifications

The specifications of Intel Loihi2 Neuromorphic Processor include the following.

• Intel Loihi 2 Neuromorphic Processor was released and announced on September 30, 2021
• Its manufacturing process is pre-production 7 nm through EUV lithography.
• This processor’s Die Size is 31 mm², and the transistor count is around 2.3 billion.
• They have 128 Neuromorphic Cores, where each neurocore can simulate 8,192 neurons. So, 128 cores x 8,192 neurons = approx x 1.05 million neurons per chip.
• Synapse Count is ~120 million, estimated but variable based on routing.
• Each core supports completely programmable microcode, like arithmetic, control flow & comparisons.
• It supports 32-bit integer payload spikes, which allows richer data transmission across the network.
• This processor supports three-factor learning rules.
• Core Memory is ~192 KiB SRAM for each core through flexible memory bank allotment or configuration.
• Every neuron model includes up to 4,096 state variables.
• Embedded x86 Cores onboard, which handle control, microcode execution, and host interfaces.
• It supports 3D scalable mesh routing, which allows solid chip-to-chip stacking & routing.
• It has standard interfaces like SPI, 10GbE, GPIO, 2500BASE KX, and 1000BASE KX to help integration through peripheral systems & sensors.
• This processor operates approximately ~100 mWatts with up to ~1 W maximum consumption.

Intel Loihi2 Neuromorphic Processor Architecture

The Intel Loihi2 Neuromorphic Processor is designed on the foundations of earlier neuromorphic systems of Intel, like Loihi & Loihi 2. Thus, it is designed to imitate the biological formation of the brain.

Key Features

The key features of the Intel Loihi2 processor architecture include the following.

Neuromorphic Design

This processor mimics the neural network of a brain in a hardware-optimized way. So it uses spiking neural networks, a group of neural networks that represent how neurons communicate within the brain by producing spikes when they meet certain conditions.

Event-Based Computing

The Loihi2 processor utilizes event-driven architecture to activate computational resources when related sensory events or inputs occur. Thus, this can considerably decrease power consumption to make it compatible with IoT and edge AI applications.

Massive Parallelism

Loihi2 uses large-scale parallelism where every neuromorphic core in this processor simulates thousands of synapses and neurons. Thus, it is very efficient to handle multi-dimensional and complex tasks. This architecture simply supports a high level of parallel execution to mimic the brain in processing data in parallel.

Low Power Consumption

Loihi2 processor achieves ultra-low power consumption while performing complex tasks. This chip consumes power only when required by processing data in an event-based manner,

Real-Time Learning

Loihi2 processor can be programmed to learn in real-time from its inputs by adapting & reconfiguring itself when it processes data. So this capability allows for dynamic adjustment and on-the-fly learning to changes within input stimuli, related to how humans are trained from experience.

Integrated Memory & Processing

Loihi2 mixes memory & processing elements in a manner that looks like the biological brain’s architecture closely. Thus, this integration assists in decreasing the energy consumption and latency associated with data transmission between separate memory & processing units.

Optimized for Sensory Data

Loihi2 processor is well-suited particularly for applications that use sensory data processing, like hearing, touch, and vision. This chip processes inputs from a variety of sensors to execute complex computations, learn to identify patterns, and make decisions depending on real-time data.

Programmable Synaptic Weights

The synaptic weight in a neuromorphic processor represents the strength between neurons. Thus, these weights are programmable, which allows for creating different neural network topologies & performances. This is important for pattern recognition, motor control, and decision-making in AI systems tasks.

Intel Loihi2 Architecture Components

The neuromorphic architecture has different components. The breakdown of this architecture is discussed below.

SNN Cores

This processor includes various SNN cores or spiking neural network cores, which simulate spiking neurons & synapses. These are event-driven, and they compute only when spikes occur. Each core in this core simulates up to ~1 million neurons for each chip. Thus, it communicates via discrete events by mimicking how biological neurons transmit data.

Programmable Neuron Models

Loihi 2 processor supports programmable neuron & synapse models. Thus, it allows researchers to modify the neurons’ behavior. This makes it flexible to a variety of brain-inspired learning rules like spike-timing dependent plasticity, or STDP.

Asynchronous Mesh Network

This mesh network helps to interconnect cores and also with packet-based NoC (network-on-chip). This mesh interconnection lets neuron cores converse spikes without depending on synchronized clocks by allowing low-latency and scalable communication.

Embedded Microcode Processors

Each core has a small RISC programmable processor, used for local computation, next learning & neuron state updates. Thus, they support implementing advanced learning algorithms directly on-chip.

Learning Engines or Plasticity Mechanisms

Loihi 2 processor has programmable plasticity engines that implement on-chip and real-time learning. It supports multiple learning rules simultaneously. which is a significant upgrade from Loihi 1.

Hierarchical Memory

Memory architecture is tightly coupled and hierarchical with the neuron cores, which store neuron states, spike routing tables, and synaptic weights. Memory access can be event-driven to decrease power consumption.

I/O & External Communication Interfaces

This processor has I/O & external communication interfaces like PCIe or Ethernet, used for integration through host systems. These can interface with actuators, sensors, and fixed AI systems, mainly for hybrid workloads.

Intel Loihi2 1 Vs Intel Loihi2 2

The difference between Intel Loihi 1 and Intel Loihi 2 processors includes the following.

Intel Loihi 1	Intel Loihi2
Loihi 1 is an advanced neuromorphic chip by Intel..	Loihi 2 is the latest neuromorphic research chip by Intel.
This processor is designed to imitate brain-like processing, allowing efficient and adaptive machine learning based applications.	This processor implements spiking neural networks through programmable dynamics, next modular connectivity & optimizations
Its architecture has neuromorphic cores – 128 and Lakemont cores -3.	Its architecture has asynchronous chip-to-chip signaling bandwidths & three-dimensional mesh network topologies.
It has 130,000 artificial neurons & 130 million synapses.	It supports neurons up to 1 million & synapses – 120 million.
This processor has fixed memory allocation for each core.	This processor has flexible memory banks, which allow dynamic allocation depending on application requirements.
Neuron programmability is configurable; however, it has fixed-function neurons.	Neuron programmability has a full instruction set for neurons, which allows for supporting diverse SNNs.
Binary-valued spike messages.	Integer-valued payload support mainly for spikes with small performance, otherwise, energy overhead.

Real World Example for Loihi2

Here are some actual applications and research projects that used Intel’s neuromorphic chips.

1. Adaptive Robot Navigation

Institution: ETH Zurich

Project: A mobile robot learned to navigate dynamic environments using on-chip reinforcement learning.

Key Advantage: Real-time, low-power navigation in changing environments without cloud support.

2. Olfactory Sensing with SNNs

Institution: Cornell University

Project: Classifying odors using neuromorphic olfactory sensors and Loihi for real-time inference.

Insight: Emulates the biological olfactory system and achieves high efficiency in detecting hazardous gases.

3. Event-Based Vision for Gesture Recognition

Institution: University of Manchester

Project: Used Dynamic Vision Sensors (DVS) with Loihi to recognize hand gestures in real time.

Benefit: Very low latency and ultra-low power compared to traditional CNNs on GPUs.

4. Edge-AI for Brain-Machine Interfaces (BMI)

Institution: Columbia University

Project: Decoding neural signals for prosthetic control.

Why Loihi? Real-time closed-loop learning and ultra-low power budget.

5. SLAM (Simultaneous Localization and Mapping)

Institution: Intel Labs

Project: SNNs implemented on Loihi for SLAM in mobile robots.

Outcome: Power-efficient navigation system that adapts to new environments quickly.

Advantages

The advantages of the Intel Loihi2 neuromorphic processor include the following.

• Loihi 2 processor uses a spiking neural network architecture to mimic how neurons within the brain communicate with spikes. Thus, it allows for very low power consumption, particularly for sensory data processing.
• Supports on-chip learning like supervised, unsupervised & reinforcement learning.
• It can adapt & learn in real-time, not like traditional neural networks, which generally need retraining.
• Loihi 2 has up to 120 million synapses and above 1 million neurons on a single chip.
• It utilizes massive parallelism for different tasks like anomaly detection, robotic control, and pattern recognition.
• Loihi 2 supports the new Lava software framework, which provides more accessible and flexible tools for developers.
• It is simple to simulate, deploy, and develop neuromorphic algorithms as compared to previous platforms.
• It is built with Intel 4 process node, which is more power & area-efficient as compared to its predecessor.
• It provides better scalability & integration potential, mainly for future AI hardware.
• It is suited for edge AI applications wherever power efficiency and low latency are critical.
• This processor processes multi-modal sensor data in real-time.
• It uses event-driven or asynchronous computation, where processing only occurs when required, which decreases energy usage.
• It supports efficient and fast inter-core communication with mesh routing, mainly for spike messages.
• Multiple Loihi 2 processors can be integrated to create larger neural networks.

Disadvantages

The disadvantages of the Intel Loihi2 neuromorphic processor include the following.

• Intel’s Lava framework has better programmability.
• It supports only a few tools, libraries, and the community.
• Neuromorphic computing needs a different programming paradigm.
• Most trendy deep learning models are not simply portable to this processor without significant re-engineering.
• Conversion is non-trivial from usual architectures to spiking models, which involve trade-offs in accuracy.
• It is still a research platform, including limited deployment within production environments.
• Industry adoption is very slow because of uncertainty regarding return on assets, unfamiliarity, and a lack of standards.
• Intel’s neuromorphic chips are not open for general purpose, but available through research partnerships and Intel’s Neuromorphic Research Community (INRC).
• It performs inefficiently or poorly on dense-data tasks like NLP compared to GPUs and large-scale image classification.
• Lack of consistent benchmarks for neural networks spiking makes it difficult to contrast performance significantly against CPUs or GPUs.

Applications

The applications of the Intel Loihi2 neuromorphic processor include the following.

• This processor mimics the neural structure of the brain with SNNs (Spiking Neural Networks).
• It performs AI tasks with very low power consumption, real-time response, and high parallelism to make it perfect for latency-sensitive and energy-constrained applications.
• This processor is used in robotics & autonomous systems for real-time motor control, adaptive robotic learning, low-latency sensorimotor integration, etc.
• It is used for event-based vision processing through DVS (dynamic vision sensors)
• This processor is used for object classification, lane detection, and gesture detection.
• It is used in Edge AI and IoT Devices for Audio and visual sensing, speech commands, and anomaly detection.
• It solves combinatorial optimization problems, such as route planning and resource allocation.
• It runs large-scale brain-inspired networks to create cognitive functions
• This processor is perfect for voice assistants, embedded NLP on edge, and Chatbots.
• It is used in Real-Time signal processing for seismic data processing & audio classification.

FAQs

1). What makes Intel Loihi 2 unique?

Loihi 2 mimics the brain’s neural architecture using spiking neural networks, enabling ultra-low-power, real-time learning and adaptive AI.

2). How is Loihi 2 different from GPUs?

Unlike GPUs that process data in dense, parallel blocks, Loihi 2 uses event-driven spikes and supports on-chip learning, making it faster and far more energy-efficient for certain AI tasks.

3). Can Loihi 2 be used for edge AI?

Answer: Yes, Loihi 2 is ideal for edge AI because it delivers high performance with minimal power consumption, perfect for real-time sensing and decision-making.

4). What software is used to program Loihi 2?

Loihi 2 is programmed using Lava, an open-source Python-based neuromorphic computing framework developed by Intel.

5). Is Intel Loihi 2 available commercially?

As of now, Loihi 2 is not widely available commercially but is accessible through Intel’s research programs and select partnerships.

Reference Links: Intel Intel Neuromorphic Research Community (INRC).

Lava source code on GitHub.

Thus, the Intel Loihi2 Neuromorphic Processor signifies a significant progression in the neuromorphic computing field. By mimicking the human brain’s function and structure, this processor handles different tasks like real-time learning, sensory processing, and decision-making with an efficiency that fixed computing systems struggle to match. Thus, this technology has long-term potential that revolutionizes many industries. Here is a question for you: What are the different types of neuromorphic processors available in the market?