Loihi

Loihi. Loihi is a neuromorphic computing research chip developed by Intel Labs, designed to mimic the architecture and function of the mammalian brain. It implements spiking neural networks using a manycore processor architecture composed of asynchronous, event-driven cores. The chip is a central platform for advancing research in artificial intelligence, neuroscience, and energy-efficient computing.

Overview

The project represents a significant departure from traditional von Neumann architecture used in conventional central processing units and graphics processing units. Inspired by biological systems, its primary goal is to achieve substantial gains in computational efficiency and learning capabilities for specific problem classes. Research utilizing the platform is often conducted in collaboration with academic institutions like Cornell University and national laboratories such as Sandia National Laboratories. The design emphasizes co-located memory and compute, avoiding the energy-intensive data movement that characterizes standard supercomputer designs.

Architecture

The physical architecture consists of a mesh of neurosynaptic cores, each containing a programmable learning engine and a set of axons and neurons. These cores communicate via asynchronous packet switching on a network-on-chip, where spikes are transmitted as messages, similar to protocols used in some parallel computing systems. Each core can simulate thousands of spiking neurons, with the entire chip containing over 130,000 neurons and 130 million synapses. The fabrication process utilizes Intel's established 14 nm process technology, and the chip's operation is governed by principles from computational neuroscience. This design allows for extreme parallelism and fine-grained event-driven processing.

Software and programming

Programming and simulation are enabled through the Intel software framework known as Lava, an open-source ecosystem for developing neuromorphic applications. The framework provides abstractions that allow researchers to compose networks without managing low-level hardware details, supporting development in Python (programming language). Key computational primitives include leaky integrate-and-fire neuron models and spike-timing-dependent plasticity learning rules. The software stack is designed to be compatible with other neuromorphic systems, fostering collaboration within the broader community, including institutions involved in the Human Brain Project. This approach allows algorithms to be tested in simulation before deployment on the physical hardware.

Research and applications

Primary research applications focus on problems involving real-time processing, adaptive control, and sparse, event-based data. Notable experiments have demonstrated efficient solutions for constraint satisfaction problems, graph search algorithms, and olfaction-inspired classification. The chip's low power consumption during specific tasks makes it a candidate for edge computing applications in robotics and autonomous vehicle systems. Collaborative studies with partners like the University of Tennessee have explored its use for scientific computing, including simulations of quantum chemistry systems. Its ability to learn continuously from data streams also aligns with research into lifelong learning for artificial intelligence.

Development and history

The project was first publicly announced by Intel in September 2017, emerging from years of internal research into post-Moore's Law computing paradigms. The second-generation chip, Loihi 2, was introduced in 2021, fabricated on a pre-production Intel 4 process node, offering improved performance and programmability. Development has been led by teams at Intel Labs in locations such as Hillsboro, Oregon and has been influenced by earlier neuromorphic efforts like IBM's TrueNorth project. The research community accesses the technology through the Intel Neuromorphic Research Community and cloud-based systems like Kapoho Bay. Future roadmaps may explore integration with other emerging technologies, including quantum computing and advanced memory technology.

Category:Intel microprocessors Category:Neuromorphic engineering Category:Artificial intelligence projects