Intel introduced its second-generation neuromorphic chip, Loihi 2, with the aim to provide tools for research in the field of neuromorphic computing. In addition, Intel has released Lava, a software framework to build neuromorphic apps both on conventional and neuromorphic hardware.
Our second-generation chip greatly improves the speed, programmability, and capacity of neuromorphic processing, broadening its usages in power and latency constrained intelligent computing applications.
The idea behind neuromorphic computing is using a processor modeled after the brain, with neurons and synapses to carry through computing tasks.
Introduced three years ago, Loihi 2's predecessor has been used by some 250 researchers to build solutions to control drones or robot arms, optimize train schedules, search databases, and to learn to identify different odors.
Loihi 2 attempts to circumvent a number of limitations of the original architecture, including its reduced flexibility, limited numerical precision, and congestion between interconnected Loihi chips, among others. This required to redesign some of circuits of the processor, which ended up including 1 million neurons, up from 125,000, in half the size using Intel 4 process, previously known as Intel 7nm.
Among the improvements in Loihi 2, it is worth to mention increased precision thanks to spike signals that carry both timing and magnitude parameters; enhanced programmability with support for arithmetic, comparison, program control flow and more, which expands the set of spiking neural network (SNN) models that can be trained. Loihi 2 also improves its learning capabilities and speed.
Alongside Loihi 2, Intel has also released Lava, an open source software framework that aims to make it easier to build neuromorphic applications by providing a set of building blocks. What makes neuromorphic programs different from traditional ones is they are strictly event-based and asynchronous, which makes them more time-dependent.
Lava applications allow neuromorphic platforms to intelligently process, learn from, and respond to real-world data with great gains in energy efficiency and speed compared to conventional computer architectures.
Lava maps to heterogeneous hardware, including non-neuromorphic platforms. The latter is a great feature for anyone wishing to experiment with neuromorphic programming on conventional CPUs/GPUs.
Inspired by Communicating Sequential Process (CSP), the basic building blocks in a Lava program are processes, which encapsulate state and behaviour. Lava's standard library provides a number of higher-level processes that implement various kinds of neuron models, neural network connection topologies, IO processes, and so on. Processed are then combined and used in algorithms, such as optimization, deep learning, and more.
You can find a short tutorial on installing and using Lava on Lava's GitHub repo.