Intel intros Loihi 2 research chip and open Lava software for neuromorphic work

Intel’s research work in neuromorphic computing has advanced to the point of a second-generation research chip called Loihi 2, bringing the promise of astounding discoveries.

Loihi 2 improves speed, the ability to program and to advance neuromorphic processing, according to a statement from Mike Davies, director of the chipmaker’s Neuromorphic Computing Lab on Thursday.

Just as important, Davies announced Intel’s Lava software framework is now open-sourced to allow cross-platform collaboration and to speed up Intel’s progress towards commercial viability of neuromorphic computing.   The framework is available with a free download on GitHub.

Neuromorphic computing is so named because it relies on neuroscience to make chips that function more like the biological brain, which means they will improve energy efficiency, computational speed and efficient learning across edge applications for vision, voice and gesture recognition.

Such applications have already been used by Intel and partners in a variety of robotics and search applications. For example, in 2020, Intel announced a robotic arm prototype to assist patients with spinal injuries in performing daily tasks.

In another example, Intel has worked with researchers in Singapore on artificial skin combined with vision sensors to help robots detect touch much faster.

 And a first gen Loihi test chip has been used to approximate the sense of smell. That work relies on research about olfactory systems in animals that looked at the electrical activity of animals as they smell odors.  Intel researchers used circuit diagrams and electrical impulses from that animal olfactory activity to build algorithms which were put on the Loihi test chip.

Loihi 2 builds on the abilities of three years of work with first-gen Loihi, offering up to 10 times faster processing and up to 1 million neurons per chip. Each Loihi 2  core measures 0.21 mm2 supporting up to 8,192 neurons while each  first gen core measures 0.41 mm2, supporting 1,024 neurons—up to 15 times greater “resource density,” in Intel parlance.

Intel worked with its tech development group to make Loihi 2 on a pre-production version of the Intel 4 process, which is partly distinguished for using extreme ultraviolet lithography.  Analysts and the financial community have been closely following Intel’s process technology advances to judge how competitive Intel will remain against global chip design and fabrication competitors.

RELATED: Intel seeks to soar again on new process and packaging roadmap

In July, Intel changed the way it names process nodes to reflect performance per watt and not just optimization of process density.  It has dropped direct references to process node size. Intel is currently in the work on a 10-nm SuperFin chip and the next chip will simply be called Intel 7, to go into production early in 2022, followed by the Intel 4 and the Intel 3.

Intel on Thursday also said its Neuromorphic Research Community has growth to 150 members which now include Ford, Georgia Institute of Technology, Southwest Research Institute and Teledyne-FLIR.

Intel is now offering two Loihi 2-based systems through the cloud to some members in the Intel NRC called Oheo Gulch and Kapoho Point.

The Loihi 2 and open Lava announcements brought some strong endorsements from various researchers in perception and control of flying drones used in package delivery, inspection and rescue efforts.

To increase usability of drones for commercial applications “it is critical to advance computer vision for on-board cameras in ways that are also computationally efficient,” said Davide Scaramuzza, a professor and director of the robotics and perception group at the University of Zurich in a statement.

Queensland University of Technology is interested in Loihi 2 for work in more complex neuronal modules to help implement biologically inspired navigation and map formation algorithms. Los Alamos National Lab is already using the first gen Loihi to research tradeoffs between quantum and neuromorphic computing and to investigate the backpropagation algorithm, considered a building block for training neural networks.